Modelling Driver Behaviour in Automotive

Environments
Carlo Cacciabue (Ed.)

Modelling Driver
Behaviourin Automotive
Environments
Critical Issues in Driver Interactions with Intelligent Transport Systems

~ Springer
P.Carlo Cacciabue,Ph.D.
EC IRC -IPSC, Italy

British Library Cataloguing in Publication Data

A catalogue record for this book is available from the British Library

Library of Congress Control Number: 2006937872

ISBN-10: 1-84628-617-4 e-ISBN-I0: 1-84628-618-2

ISBN-13: 978-1-84628-617-9 e-ISBN-13: 978-1-84628-618-6

Printed on acid-free paper

Springer-Verlag London Limited 2007

Whilst we have made considerable efforts to contact all holders of copyright material contained in this
book, we may have failed to locate some of them. Should holders wish to contact the Publisher, we will
be happy to come to some arrangement with them.

Apart from any fair dealing for the purposes of research or private study, or criticism or review, as
permitted under the Copyright, Designs and Patents Act 1988, this publication may only be reproduced,
stored or transmitted, in any form or by any means, with the prior permission in writing of the
publishers, or in the case of reprographic reproduction in accordance with the terms of licences issued
by the Copyright Licensing Agency. Enquiries concerning reproduction outside those terms should be
sent to the publishers.

The use of registered names, trademarks, etc. in this publication does not imply, even in the absence of
a specific statement, that such names are exempt from the relevant laws and regulations and therefore
free for general use.

The publisher makes no representation, express or implied, with regard to the accuracy of the infor-
mation contained in this book and cannot accept any legal responsibility or liability for any errors or
omissions that may be made.

9 8 7 6 543 2 1

Springer Science+Business Media

springer. com
Contents

Editorial .................................................................................. viii

List of Contributors. ................................................................... XII

Chapter 1. International Projects and Actions on Driver Modelling

1. Modelling Driver Behaviour in ED and International Projects. .......... 3

Maria Panou, Evangelos Bekiaris and Vassilis Papakostopoulos

2. TRB Workshop on Driver Models: A Step Towards

a Comprehensive Model of Driving? .......................................... 26
Delphine Cody and Timothy Gordon

3. Towards Monitoring and Modelling for Situation-Adaptive

Driver Assist Systems. ........................................................... 43
Toshiyuki Inagaki

Chapter 2. Conceptual Framework and Modelling Architectures 59

4. A General Conceptual Framework for Modelling Behavioural

Effects of Driver Support Functions. .......................................... 61
Johan Engstrom and Erik Hollnagel

5. Modelling the Driver in Control................................................ 85

Bjorn Peters and Lena Nilsson

6. From Driver Models to Modelling the Driver: What Do

We Really Need to Know About the Driver? 105
Oliver Carsten

v
VI Contents

Chapter 3. Learning and Behavioural Adaptation 121

7. Subject Testing for Evaluation of Driver Information Systems

and Driver Assistance Systems - Learning Effects and
Methodological Solutions................................................... .... 123
Klaus Bengler

8. Modelling Driver's Risk Taking Behaviour.................................. 135

Wiel Janssen

9. Dealing with Behavioural Adaptations to Advanced Driver

Support Systems. .................................................................. 147
Farida Saad

Chapter 4. Modelling Motivation and Psychological Mechanisms 163

10. Motivational Determinants of Control in the Driving Task................ 165

Ray Fuller

11. Towards Understanding Motivational and Emotional Factors

in Driver Behaviour: Comfort Through Satisficing......................... 189
Heikki Summala

12. Modelling Driver Behaviour on Basis of Emotions and Feelings:

Intelligent Transport Systems and Behavioural Adaptations.............. 208
Truls Vaa

Chapter 5. Modelling Risk and Errors 233

13. Time-Related Measures for Modelling Risk in Driver Behaviour....... 235

Richard van der Horst

14. Situation Awareness and Driving: A Cognitive Model..................... 253

Martin Baumann and Josef F. Krems

15. Driver Error and Crashes ........................................................ 266

Dianne Parker

Chapter 6. Control Theory Models of Driver Behaviour 275

16. Control Theory Models of the Driver......................................... 277

Thomas Jiirgensohn

17. Review of Control Theory Models for Directional

and Speed Control................................................................. 293
David H. Weir and Kevin C. Chao
 Contents VB

Chapter 7. Simulation of Driver Behaviour 313

18. Cognitive Modelling and Computational Simulation

of Drivers Mental Activities. .................................................... 315
Thierry Bellet, Beatrice Bailly, Pierre Mayenobe and Olivier Georgeon

19. Simple Simulation of Driver Performance for Prediction

and Design Analysis.............................................................. 344
P. Carlo Cacciabue, Cristina Re and Luigi Macchi

Chapter 8. Simulation of Traffic and Real Situations 377

20. Real-Time Traffic and Environment Risk Estimation

for the Optimisation of Human-Machine Interaction...................... 379
Angelos Amditis, Aris Polychronopoulos and Evangelos Bekiaris

21. Present and Future of Simulation Traffic Models.. ......................... 400

Fabio Tango, Roberto Montanari and Stefano Mariani

Index...................................................................................... 429
Editorial

The implementation of information technology and automation has been the

driving force of the development of technology in the last decades. At the same
time the presence of humans in control of systems has been kept as the locus of
design principles. Only more recently, fully unmanned technology is beginning to
find its application, in a limited number of domains, such as urban guided trans-
port systems and military aviation. However, even in the cases of totally automatic
systems, it is not possible to avoid the assessment of the human-in-control prin-
ciple, as the overall remote operator of the fully automatic systems remains to
be accounted for in the design and development processes. For these reasons, the
consideration of the user and controller of technologically advanced systems is
one of the most relevant issues for design development, and during production and
implementation processes.
The use of "intelligent" systems implies that the level of autonomy and the
possibility to delegate control processes to technology and automation has im-
proved enormously. The human being has been freed from performing a number
of activities and has progressively been removed from the direct control loop of
systems, in favour of high level decision making processes.
Therefore, in parallel to the development of technology, the need to account for
the behaviour of the human being has progressively evolved from the consideration
of the human-manual-controller to the human-supervisor of processes, procedures
and performances of automatic control systems. The evaluation of behavioural
performance has been replaced by the analysis of cognitive and mental processes.
In other words, the demand for modelling manual and behavioural activities has
been replaced and combined with the need for modelling cognition. This is the
requirement that has mostly affected the development of new technologies and
interfaces in modem control devices.
This feature of design and development of technological systems is common to
many different domains, from energy production, chemical and process industry,
transport and health care. In this scenario, the automotive transport is the domain
mostly affected by the need to consider the multiplicity of human behaviour,
as it presents the highest possible variety of operating environments, of human

viii
 Editorial ix

behaviours, and it offers many different technological solutions for all different
control processes.
In reality, many automation controls currently applied to vehicles already
contain models of a certain complexity of cognition and behaviour, based mostly
on dynamic manifestation of control operations. The "automatic gearbox" of cer-
tain vehicles is a typical example of this type of control systems, which adapt
dynamically and independently to different "driving styles", measured through
intrinsic evaluation of behavioural variables, such as rate of accelerator pressure,
overall speed, etc. Another example is the system that manages the availability
of in vehicle information systems (IVIS), such as telephones or radios. In this
case, certain IVIS managers adapt to the environmental situations, by inhibiting
or discouraging the use of certain IVIS in risky situations.
The models of cognition and behaviour that are implemented in such types of
vehicle control systems are naturally elementary from the cognitive point of view.
However, this shows that this industrial field, both in terms of vehicle integrators
and original equipment manufacturers, needs to apply at design level and integrate
at implementation level adequate models of driver behaviour. These models are
equally important for academic and research purposes, where more complex and
varied solutions can be proposed and studied in relation to theoretical paradigms of
different nature and targets. Another area where modelling of driver behaviour is
essential is the transport safety authorities and regulators, where the consideration
of driver performance becomes essential in setting standards and rules governing
new and future regulations of vehicle control systems, road infrastructures and
traffic management. Similarly, models of drivers are necessary for the study of
accidents and investigation of root causes.
The availability of models and paradigms of driver behaviour at different levels
of complexity and development is therefore quite obvious according to the field of
application.
This book offers to the reader the possibility to assess different approaches and
considerations in relation to driver behaviour modelling, resulting from different
fields. Indeed, the authors of the different manuscripts come from the industrial
area, both car and original equipment manufactures and integrators, from the re-
search and academic fields and from national and international regulators and
automotive transport authorities.
More in detail, Chapter 1presents the ongoing activities in International Projects
and Actions on Driver Modelling. In particular, the European sustained research
Projects carried out over the last decades and presently under development are
reviewed in the paper by Panou, Bekiaris, and Papakostopoulos. Similarly, the US
research actions on driver models and a recently held workshop on these issues
are discussed in the paper by Cody and Gordon. The last paper of this Chapter,
by Inagaky, also revises the actions in Japan on driver modelling, focusing on
monitoring and modelling situation-adaptive driver assistance systems.
After this initial review, more specific subjects are dealt with, beginning with
the existing Conceptual Frameworks and Modelling Architectures (Chapter 2)
that sustain the development of specific models of driver behaviour. In all three
x Editorial

papers, a short historical review of paradigms and architectures for considering

the human element in a Driver-Vehicle-Environment perspective is performed. The
three papers then focus on a specific modelling architecture that enables the reader
to consider very high level structures of interaction, primarily in the papers by
Engstrom and Hollnagel, and by Peters and Nilsson, and more practical and field
focused architectures in the paper by Carsten.
Chapter 3, contains a overall discussion and overview of one of the major issues
that affect driver behaviour modelling: Learning and Behavioural Adaptation. This
issue is dealt with from different perspectives, beginning with the view on Testing
for Evaluation of driver information systems and driver assistance systems, by
Bengler. The crucial issue of risk taking and risk perception is discussed in the
work of Janssen. The Chapter is then completed by the paper of Saad, that offers a
wide and distributed review of theoretical stands on adaptation and processes that
affect driver performance.
In Chapter 4 another essential and widely debated aspect of driver behaviour
is tackled: Modelling Motivation and Psychological Mechanisms. The three con-
tributors to this Chapter offer different and, in some cases, controversial views on
specific formulations and algorithms that can be applied to describe and account for
these critical factors governing driver behaviour. In particular, Fuller concentrates
on determinants of control in the driving task. Summala describes motivational
and emotional factors through the concept of "satisfying". Vaa offers an overview
of his long standing arguments on emotions and feelings.
Chapter 5 deals with Modelling Risk and Errors. The issue of risk management
and risk perception, already dealt with in the Chapter about adaptation, is further
developed here in the paper by Van der Horst, where the specific problem of time-
related measures for modelling risk are discussed in detail. The other two papers
focus more closely on aspects associated with human error. The paper of Baumann
and Krems is dedicated to critical causes of human error. A model based on the
evaluation of Situation Awareness for different driving situations is developed. The
paper by Parker offers instead a wider overview and discussion on the human error
making more correlated to classical and well established theories. The specific
application to the automotive domain demonstrates the possibility to apply such
theories to this domain as it has been done in other cases, e.g., aviation, nuclear
energy etc.
Chapter 6, introduces the second part of the book, dedicated to the review of a
number of numerical simulation and computerised implementation of the theories
and paradigms described in the previous five chapters of the book. In particular,
Chapter 6 discusses the Control Theory Models ofDriver Behaviour. The paper by
Jiirgensohn offers a very comprehensive review of many Control Theory Models
of the Driver that have been developed in the past and are still nowadays very
valuable numerical approaches for describing human machine interactions. The
paper by Weir and Chao also offers a review of control theory approaches, but
then quite readily focuses on a specific control theory approach developed by the
authors over a number of years for describing Directional and Speed Control over
a certain period of time.
 Editorial xi

Chapter 7, concentrates on the problem of the overall Simulation of Driver

Behaviour and representation of the interaction with the vehicle and environment.
The paper by Bellet, Bailly, Mayenobe, and Georgeon describes a computational
simulation of drivers mental activities called COSMODRIVE, based on several
Artificial Intelligence approaches such as the Blackboard architecture and spe-
cific simulation languages that enable the fast and simple description of driver
performance. The paper by Cacciabue, Re and Macchi offers a similar simula-
tion approach, called SSDRIVE, which describes the driver behaviour by means
of Object Oriented Languages and applies the theoretical paradigm described in
Chapter 2 of the book and exploits the power of control theory to describe driver
control actions.
Finally, Chapter 8 deals with the issue of Simulation of Traffic and Real Situa-
tions. The need to simulate the driving context is directly associated to the ability of
modem simulations to predict behaviour and driver interactions with the vehicle in
many driving situations. Consequently in order to make the simulation as realistic
as possible, it is necessary that the models that are coupled to driver simulation are
equally detailed and representative of real traffic situations. Two papers complete
this chapter. The paper by Amditis, Polychronopoulos, and Bekiaris concentrates
on real-time traffic and environment risk estimation, while the paper by Tango,
Montanari, and Marzani describes a set of existing simulations of traffic and dis-
cusses the new orientation and perspectives of future models and simulations of
traffic.
The above description of the Chapters demonstrates that the Book can be read as
a whole in order to get a general perspective of the ongoing topics of development
for present generation of automation and intelligent support systems. At the same
time, it offers the overview on open issues characterising research into new areas of
concern for future control perspectives. However, it is equally possible to consider
selected readings of papers and chapters, when the needs of the reader focus on well
defined and specific subjects, such as for example driver adaptation or simulation
of DVE interactions.
The quality and completeness of this Book rests primarily on the excellence of
the papers that are included. For this reason, the editor of this book is deeply grateful
to all authors who have diligently, professionally and proficiently collaborated to
its development. The editor has simply acted as catalyser and integrator of ideas
and competences, ensuring that the pieces of this puzzle could come together in a
consistent and coherent vision of the problems under scrutiny. The success of this
endeavour rests eventually on the quality of its content and on the response of the
scientific community to which it is addressed.

P. Carlo Cacciabue
List of Contributors

Angelos Amditis Oliver Carsten

I-SENSE Group, University of Leeds,
Institute of Communications and Institute for Transport Studies,
Computer Systems, Leeds LS2 9JT, UK.
Athens, Greece.
Kevin C. Chao
Beatrice Bailly Dynamic Research, Inc.,
INRETS- LESCOT, Torrance, CA, USA.
Bron cedex, France.

Martin Baumann Delphine Cody

Chemnitz University of Technology, UC Berkeley,
Department of Psychology, California PATH,
Chemnitz, Germany. Richmond, CA, USA.

Evangelos Bekiaris Johan Engstrom

Hellenic Institute of Transport, Volvo Technology Corporation,
Athens, Greece. Goteborg, Sweden.

Thierry Bellet Ray Fuller

INRETS-LESCOT, Trinity College Dublin,
Bron cedex, France. Department of Psychology,
Dublin 2, Ireland.
Klaus Bengler
BMW Group,
Forschung und Technik, Olivier Georgeon
Munich, Germany. INRETS-LESCOT,
Bron cedex, France.
P. Carlo Cacciabue
EC, Joint Research Centre, Timothy Gordon
Institute for the Protection and University of Michigan,
Security of the Citizen, Transportation Research Institute,
Ispra (VA), Italy. Ann Arbor, MI, USA.

xii
 List of Contributors Xl11

Erik Hollnagel Roberto Montanari

Ecole des Mines de Paris, University of Modena and Reggio
Pole Cindyniques, Emilia,
Sophia Antipolis Cedex, France. Dipartimento di Scienze e Metodi
dell'lngegneria,
Reggio Emilia, Italy.
Toshiyuki Inagaki
University of Tsukuba,
Department of Risk Engineering, Lena Nilsson
Tsukuba, Japan. VTI,
Linkoping, Sweden.
WielJanssen
TNO Defence, Security & Safety,
MariaPanou
BU Human Factors,
Hellenic Institute of Transport,
Soesterberg, The Netherlands.
Athens, Greece.

Thomas Jiirgensohn
HFC Human-Factors-Consult Vassilis Papakostopoulos
GmbH, Hellenic Institute of Transport,
Berlin, Germany. Athens, Greece.

DianneParker
JosefF.Krems University of Manchester,
Chemnitz University of Technology, School of Psychological
Department of Psychology, Sciences,
Chemnitz, Germany. Manchester M13 9PL, UK.

LuigiMacchi
BjornPeters
EC, Joint Research Centre,
VTI,
Institute for the Protection and
Linkoping, Sweden.
Security of the Citizen,
Ispra (VA), Italy.
Aris Polychronopoulos
Stefano Marzani I-SENSE Group,
University of Modena and Reggio Institute of Communications and
Emilia, Computer Systems,
Dipartimento di Scienze e Metodi Athens, Greece.
dell'lngegneria,
Reggio Emilia, Italy.
Cristina Re
EC, Joint Research Centre,
PierreMayenobe Institute for the Protection and
INRETS-LESCOT, Security of the Citizen,
Bron cedex, France. Ispra (VA), Italy.
xiv List of Contributors

Farida Saad Truls Vaa

INRETS-GARIG, Institute of Transport Economics
Champs-sur-Marne, France. (T0I),
OSLO, Norway.
HeikkiSummala
Richard van der Horst
University of Helsinki, Department of
TNO Defence, Security & Safety,
Psychology,
BU Human Factors,
Helsinki, Finland.
Soesterberg, The Netherlands.

FabioTango DavidH. Weir

Centro Ricerche Fiat (C.R.F.) Dynamic Research, Inc.,
Orbassano (TO), Italy. Torrance, CA, USA.
 I
International Projects and Actions on
Driver Modelling
1
Modelling Driver Behaviour in
European Union and International
Projects
M. PANOU, E. BEKIARIS AND V. PAPAKOSTOPOULOS

1.1 Introduction
Human (or operator) modelling has been an extensive area of research in many
application areas, such as artificial intelligence, aviation, probabilistic risk as-
sessments, system safety analysis and human performances in working contexts
(Cacciabue et al., 1993; Baron et al., 1980). Still, human behaviour is fairly con-
textual and substantially different from one person to another. Thus, the initial
linear models have been gradually replaced by nonlinear and even probabilistic
models, based upon artificial intelligence (AI) principles, such as artificial neural
networks or genetic algorithms. This becomes even more intrigued if we consider
a complex behavioural task such as vehicle driving.
The traffic system as a whole can be seen as being composed of three interactive
parts: vehicles, road users and the road environment. Any traffic situation is the
result of the interaction between these three systems. Normally, the traffic situation
develops as planned, but, in certain circumstances the resulting interaction will
result in a critical situation or in a crash.
The driver is a critical component of the traffic system. Attempts have been
made to estimate the importance of the driver as an accident cause (Evans, 1985).
It has been estimated that road user factors are the sole or contributory factors in
a great majority of road crashes.
There is no generally accepted model of the complete driving task. There are
detailed descriptions focusing on perception and handling aspects and reporting
what drivers really do in every possible ('normal') situation from the beginning
to the end of a journey (see McKnight and Adams, 1970). There are also more
analytical approaches focusing on driver behaviour in relation to task demands,
with the purpose of trying to explain and understand the psychological mechanisms
underlying human behaviour (Rasmussen, 1984; Michon, 1985).
Usually, car driving is described as a task containing three different levels of
demands. At the strategic level, the general planning of a journey is handled. For ex-
ample, the driver chooses the route and transportation mode and evaluates resulting
costs and time consumption. At the tactical level, the driver has to exercise ma-
noeuvres, allowing him/her to negotiate the 'right now' prevailing circumstances,

3
4 Panou, Bekiaris and Papakostopoulos

for instance, turning at an intersection or accepting a gap. Finally, at the control

(stabilisation) level the driver has to execute simple (automatic) action patterns,
which together form a manoeuvre, for example, changing the gear and turning the
wheel.
The demands imposed on the driver are met through his/her driving behaviour.
Also, the performance of the driving task is usually assigned to three different
levels: knowledge-based, rule-based and skill-based behaviour. Skill-based be-
haviour is described as data-driven, meaning that skills are performed without
conscious control and use of attention resources. They are immediate and effi-
cient. Rule-based behaviour on the other hand occurs under conscious control and
requires attention. Therefore, it is less immediate and efficient. Knowledge-based
behaviour involves problem solving and is relevant when it is not given how to act
in a specific situation. Thus, an important aspect of knowledge-based behaviour is
that reasoning is required.

1.2 Evaluation of Driver Behaviour Models

Analysing the driving task requires consideration of the dynamic interaction be-
tween drivers and the traffic system. Driver-specific factors include performance
aspects, individual dispositions and transient driver states. Driver behaviour mod-
els attempt to formalise the complex relation between the driver and the traffic
system.

1.2.1 Michon's Hierarchical Control Model

Michon (1985) proposed a simple two-way classification of driver behaviour mod-
els: One dimension distinguished between behaviour, i.e., input-output-oriented
models and internal-state-oriented models. The second dimension differentiates
between functional models and taxonomic models, where model components do
or do not interact, respectively.
According to Michon (1985), all models lack in one or more respect: they are
generally bottom-up controlled, internal models. Corresponding top-down pro-
cesses are hardly specified or they tend to be too simplistic. Michon regards cogni-
tive process models as the most encouraging step towards a valid model of driver
behaviour because these types of models combine elements of driving task analy-
sis with an information-processing approach. Therefore his Hierarchical Control
Model subdivides the driving task into three coupled and hierarchically ordered
levels, namely the strategic, the manoeuvring and the control levels. Adapting this
model, to incorporate the GADGET fourth level (see next section), one more level
is added, i.e., the 'behaviour level' (Fig. 1.1).
The strategical level includes trip planning, route choice and other general
principles including time constraints. This level is little involved in actual driv-
ing. However, it sets criteria for factors at the lower levels, like speed control
and associated subjective risk levels. At the manoeuvring level, drivers interact
 1. Modelling DriverBehaviour in ED and International Projects 5

Way of living Infinite

General plans Long

Environmental Controlled
input sees
action patterns

Environmental Automatic action 1

input 1000 sees
patterns

FIGURE 1.1. The hierarchical structure of the driving task (adapted from Michon, 1985).

with the traffic system. The control level, finally, refers to basal car control
processes. Although a dynamic relationship between the concurrent activities
is assumed, the different levels require different types of information: The
strategical level is mainly top-down (knowledge) controlled. The manoeuvring
and control levels require in addition bottom-up (data) input from the traffic
environment.
Closely associated to Michon's hierarchical model of driver behaviour is Ras-
mussen's division of operative behaviour into three levels: skill-based behaviour
refers to automatic procedures, rule-based behaviour to application of learned
rules and knowledge-based behaviour to conscious problem solving (Rasmussen,
1984). Skill-based behaviour is applied in Michon's model mainly at the control
level in the form of automatic action patterns. Ranney (1994) relates Michon's
control hierarchy to Rasmussen's taxonomy of operative behaviour. Skill-based
behaviour is applied in all familiar situations. Rule-based behaviour dominates
during standard interactions with other road users as well as in some rare situa-
tions like driving a new car, where automatic routines have to be transferred to a
new system. Knowledge is applied when driving in unfamiliar traffic networks,
in difficult environmental conditions or when skills are not fully developed as in
novice drivers.

1.2.2 The GADGET-Matrix: Integrating Hierarchical

Control Models and Motivational Models ofDriver
Behaviour
Motivational models of driver behaviour 'propose a general compensatory mecha-
nism whereby drivers adjust their driving (e.g. speed) to establish a balance between
what happens on the road and their level of acceptable subjective risk (Ranney,
1994). An important assumption of motivational models is that drivers establish a
6 Panou, Bekiaris and Papakostopoulos

constant level of risk by activating risk-compensation mechanisms when a subjec-

tive threshold is exceeded (e.g., Summala, 1988). The opportunity to compensate
risks by adjusting subjective risk levels indicates that drivers' personal motives
are as well a crucial factor for safe driver behaviour. For this reason, a fourth level
corresponding to individual dispositions has been added to hierarchical control
models of driver behaviour within the European project GADGET (Christ et aI.,
2000). The new level refers to personal preconditions and ambitions in life, and
as such has the highest priority inside the matrix because such dispositions heav-
ily influence driving decisions at lower levels. The four levels of the so-called
GADGET-Matrix are as follows (Table 1.1):

(a) Goals for life and skills for living: An individual driver's attitudes, lifestyle,
social background, gender, age and other personal preconditions that might
influence driving behaviour and accident involvement.
(b) Driving goals and context: Strategical planning of a trip; the focus is on why,
where, when and with whom one is driving.
(c) Mastery of traffic situations: Actual driving in a given context, resembles
Michon's manoeuvring level.
(d) Vehicle manoeuvring: Overlaps despite a different terminology with Michon's
car control level. The focus is on the vehicle, its construction and how it is
operated.

A safe driver has, however, not only developed skills but also knowledge about
his/her own abilities, preconditions and limits. Experienced drivers have, in addi-
tion, cognitive driving skills, such as anticipation and risk perception. In order to
cover these higher-order aspects of driver behaviour, vertical columns are added to
the so far horizontal structure of hierarchical control models (see Table 1.1). The
columns of the GADGET-Matrix are as follows:

(a) Knowledge and skills: Routines and information required for driving under
normal circumstances.
(b) Risk-increasing factors: Aspects of traffic and life associated with higher risk.
(c) Self-assessment: How good the driver reflects his/her own driving skills and
motivations.

Levels and cells of the GADGET-Matrix are not mutually exclusive - there is
large vertical as well as horizontal overlap due to the complex and cyclic nature
of the driving task, where subtasks usually have to be carried out in parallel at
different levels (e.g., routing, turning left, gap acceptance, speed control, steering,
braking, etc.).

1.2.3 DRIVABILITY Model

The most recent evaluation in driver modelling concerns the notion that driver
behaviour is not necessarily static, but evolves dynamically with time, as well
as is context-related. It is subjected not only to permanent but also temporary
 1. Modelling Driver Behaviourin ED and International Projects 7

TABLE 1.1. The GADGET-matrix (adaptedfrom Hatakka et aI., 1999).

Knowledge and skills Risk-increasing factors Self-assessment
Awareness about Risky tendencies like Awareness of
relation between personal acceptance of risks impulse control
tendencies and driving high level of sensation risky tendencies
skills seeking dangerous motives
lifestyle/life situation complying to social risky habits
peer group norms pressure
motives use of alcohol and drugs
personal values

Awareness about Risks associated with Awareness of

effects ofjourney goals physical condition personal planning skills
planning and choosing (fitness, arousal, alcohol, typical driving goals
routes etc.) alternative transport
effects of social pressure purpose of driving modes
by passengers inside the driving environment
car (rural/urban/highway)
social context and
company

Knowledge about Risks associated with Awareness of

traffic regulations wrong expectations strong and weak points of
traffic signs vulnerable road users manoeuvring skills
anticipation violations subjective risk level
communication information overload subjective safety margins
safety margins unusual conditions
inexperience

Skills concerning Risks associated with Awareness of

control of direction and insufficient skills strong and weak points of
position environmental conditions car control skills
vehicle properties (weather, friction, etc.)
physical phenomena car condition (tyres,
engine, etc.)

contributors, which mayor may not be independent. The DRIVABILITY model

(Bekiaris, Amditis, Panou, 2003) introduced as most important the following ones:

1. Individual resources, namely physical, social, psychological and mental con-

ditions of the specific driver. Physical conditions include motor, sensoric and
coordination functions. Mental status depends also on the actual level of stress,
concentration to the task and vigilance level.
2. Knowledge/skills level: This refers not only to actual driver training and ex-
perience, but also to generic knowledge, as basic education greatly influences
motivations and behaviour of the driver. This level also considers the self-
awareness of the own skills and it includes all the four levels of the GADGET
model.
8 Panou, Bekiaris and Papakostopoulos

3. Environmental factors: This includes the vehicle status, the existence of traffic
hazards, the weather, road and traffic conditions. The combinations of these
may generate a risky situation , which certainly influences DRIVABILITY.
4. Two common denominators between driver resources and environmental status,
namely workload and riskawareness.
The two intermediate factors between driver resources and the environment,
namely workload and risk assessment, are among the key issues in order to under-
stand and analyse driving performance. Risk awareness depends on three major
contributors:
1. Risk perception, namely the ability to understand/recognise the specific risk at
the specific time moment.
2. Level of attention, the ability to spot the risk in time.
3. Possible external support so as to spot the risk in time, i.e., by advanced driver
assistance systems (ADAS).

In contrast to the risk awareness level, which is rather discrete and may change
arbitrarily, the other factor, workload, is continuous and evolves with time. Even
temporary input, i.e., use of mobile phone, may have high impact on workload
for limited time periods . The major contributors to DRIVABILITY are depicted
in Fig. 1.2.

FIGURE 1.2. Contributors to DRIVABILITY.

 1. Modelling Driver Behaviour in ED and International Projects 9

The contributors shown in Fig. 1.2 are combined in a mathematical formula,

which comprises the DRIVABILITY index (DI) of each individual driver at any
given moment in time. Considering that the individual resources are the most
significant contributor, the knowledge/skills and workload are of equal importance,
while environment and risk awareness are third in importance, the overall DI is
calculated through the following empirical formula:
KSI WI EFI + RAI
DI=IRlx - x - x ----
226
where IRI is the individual resources index, KSI the knowledge/skills index, WI the
workload index, EFI the environmental factors index and RAI the risk awareness
index.
The DRIVABILITY model contributors validity are being proved within the
AIDE IP, where five modules are being developed, as components of one model,
that monitor whether the driver is engaged and/or distracted by a secondary task
and his/her availability/unavailability (related to the contributor V of Fig. 1.2),
his/her inattention/fatigue (related to contributor IV of Fig. 1.2), his/her personal
characteristics (related to the contributors I and II of Fig. 1.2) as well as the moni-
toring of the traffic and environment (related to the contributor III of Fig. 1.2). In a
few words, the Cockpit Activity Assessment (CAA) module in AIDE is intended
to monitor the activities of the driver to detect workload by visual distraction,
cognitive distraction, and signs of lateral manoeuvring intent. Furthermore, the
Driver Availability Estimator (DAE) module aims to assess the driver's 'level of
availability/unavailability' to receive and process information, according to the
requirements of the primary driving task (depending on the nature of the road in-
frastructure, the goal followed at this time, the current driving actions carried out,
etc.). The Driver State Degradation (DSD) module intends to detect and to diagnose
in real-time the driver hypo-vigilance state due to drowsiness and sleepiness situa-
tions, giving an indication about the driver's ability to execute the driving task. The
Driver Characteristics (DC) module personalises the warning and/or information
provision media, timing and intensity according to driver's profile (experience,
reaction time, average headway, etc.), explicit and implicit preferences. The Traf-
fic and Environment Risk Assessment (TERA) monitors and measures activities
outside the vehicle in order to assess the external contributors to the environmental
and traffic context and also to predict the driver's intention for lateral manoeuvre
(Boverie et al., 2005).

1.3 Driver Behaviour Adaptation Models and Their

Relation to ADAS
ADAS are currently being developed and installed within vehicles at an increasing
rate. These systems aim to improve driving safety by automating aspects of the
driving process, through information, warnings and support to the driver, hence
reducing driver workload.
10 Panou, Bekiaris and Papakostopoulos

TABLE 1.2. Driver behaviour issues when introducing ACC (Bekiaris et al., 2001).
Short term Long term
Mistrust: distrusting the ACC system Spare capacity: using spare capacity for other
in-vehicle tasks
Over-reliance: relying too much on the ACC system
Brake pedal forces: increasing brake pedal Fatigue: ACC could take over too many
forces driving tasks causing fatigue
Imitation: unequipped vehicles imitate Quick approach to vehicle in front: the
equipped vehicles development of new behaviour
Reliance on vehicle in front: vehicle in front Time-headway: driving with smaller
might have poor driving behaviour time-headways
Indication for overtaking: use ACC as an
indication of when to overtake
Overtaking: difficulties with overtaking and being overtaken

Automation can reduce driver workload in areas of decision choice, informa-

tion acquisition and information analysis. This reduction in workload should then
reduce driver error and stress, thereby increasing road safety. However potential
problems exist with the introduction of automated systems. The reduction of men-
tal workload may not always occur under conditions of system failure or when a
user is unfamiliar with the system. Under these conditions, workload may in fact
increase rather than reduce (Stokes et al., 1990). Also, changes in driver skills,
learning and behaviour, which may occur due to the shift in locus of control, may
prove detrimental and therefore predictions as to how drivers will react when lo-
cus of control shifts between the driver and the vehicle are required. Short- and
long-term driver behavioural changes with the use of an advanced cruise control
(ACC) system, from six research projects are summarised in Table 1.2 (Bekiaris
et al., 2001).
The introduction of ADAS, as with any changes to the driving environment,
may lead to changes in driver behaviour. However, the nature of these behaviour
changes in response to changes in the driving environment and has on occasions
proved to be the opposite of that which was intended. Grayson (1996) pointed out
that 'people can respond to innovation and change in ways that are unexpected,
unpredictable, or even wilfully perverse' . For example, Adams (1985) claimed that
the introduction of seat belts in vehicles leads to a perception of greater safety, in
tum leading to drivers increasing their speed on the road.
It has been suggested that improved safety cannot be predicted directly from
the efficiency resulting from improved technology, as people adapt to some kinds
of improved efficiency by taking more risks (Howarth, 1993). The introduction
of safety measures may lead to compensatory behaviours that may reduce the
benefits of the measures being implemented. This phenomenon has most recently
been described as 'behavioural adaptation' (OECD, 1990). However, previous
models explaining the behaviour have termed it as 'risk compensation' and 'risk
homeostasis' .
The most important relevant theories and issues related to driver behavioural
adaptation because of an external stimuli (ADAS introduction in our case) are
summarised below (Bekiaris etaI., 2001) ..
 1. Modelling Driver Behaviour in ED and International Projects 11

1.3.1 Automaticity
Automation refers to the mechanical or electrical accomplishment of work. Some
of the automatic components provided by ADAS act as a substitution for tasks that
humans would otherwise be capable of performing. In other cases, ADAS provides
automatic components, which carry out additional tasks that humans would not
have been capable of but will also assist in the overall driving task.
Introducing automation into the driving task offers several potential advantages
in aspects of both efficiency and safety. For example, the use of dynamic route
guidance systems will assist drivers in taking the most cost-effective route, in
terms of fuel and time, for current traffic conditions. Furthermore, automation can
reduce driver workload in areas of decision choice, information acquisition and
information analysis. This reduction in workload should then reduce driver error
and stress, thus increasing road safety.
Overtrust on a system also brings problems. Drivers may become complacent
and may not detect when a system fails. Drivers are left with a false sense of
security, thereby failing to monitor the system leading to the added disadvantage
of loosing system awareness. Drivers may also lose the opportunity to learn and
retain driving skills. Furthermore, the role of the driver may be reduced to such an
extent that their manual driving skills may degrade. This concept has been termed
'out-of-the-loop familiarity'.
Bainbridge (1987) discusses what she terms the' ironies of automation', which
occur with the changing role of the human in the human-machine relationship
when a system is automated. She points out that the more advanced a system is,
the more crucial the contributions of the human operator become. Automation aims
to eliminate the human factor; however, ironically, the human operator is required
to carry out those tasks that cannot be automated. The human operator is therefore
required to monitor the system and to take over and stabilise the system manually
in situations of system failure. However, as previously discussed manual control
skills deteriorate when using an automated system, leading an experienced user to
become inexperienced.
Bainbridge points out the loss of cognitive skills that an operator using an auto-
mated system, such as ADAS, is likely to suffer. As the retrieval of knowledge from
the long-term memory is dependent on the frequency of use, operators will loose
the benefit of long-term knowledge concerning processes. This practical knowl-
edge can be used to generate strategies in emergency or unusual situations. It is
difficult to teach practical knowledge without experience; it is thus of great concern
that when automating a system this practical experience and the reinforcement by
frequency of use will be taken away from the operator.

1.3.2 Locus of Control

Locus of control is the location of control over a situation or system. The imple-
mentation of ADAS will, in many cases, move the locus of control away from the
driver and instead will lie with the vehicle. It is therefore necessary to consider
12 Panou, Bekiaris and Papakostopoulos

the impact of this change in locus of control both on driver behaviour and also in
terms of safety. The level of control left to the driver must be carefully considered
(e.g., will drivers be given the opportunity to override vehicle decisions and vice
versa?).
Once level of control is decided upon, it is still important to consider the likely
consequences of implementation. Several concerns exist, including those discussed
in the previous section. First, drivers may feel mistrust in the system and experience
problems in handing over control to the vehicle; this factor is dependent on the
confidence a driver has in the capabilities of the system. Secondly, drivers may
overtrust the system; drivers may become dependent on the system. This may be
problematic both in cases where the system fails; the driver may not detect failure
due to reduced monitoring. Also, in situations when the driver is handed back
control, driver's skills and learning may have been diminished due to out-of-the-
loop familiarity.
Finally, driver behaviour and safety during the handover of control between the
driver and the vehicle needs to be considered. De Vos et al. (1997) investigated
safety and performance when transferring control of the vehicle between the driver
and an Automatic Vehicle Guidance (AVG) system. Drivers were found to be able
to leave the automated lane even when high-speed differences and traffic-density
differences between lanes were present. Unsafe interactions were observed in the
scenario of a low-speed manual lane. As expected, increased trust in the reliability
of the system increased driver comfort. However, as headway decreased drivers
were observed to experience greater discomfort, implying that total trust in the
system did not exist.

1.3.3 Risk Homeostasis

One of the most considered and debated models in the area is Wilde's risk home-
ostasis theory. The model bears similarities to earlier models such as that of Taylor,
and Cownie and Calderwood's. Taylor's risk-speed compensation model (1964)
claims simply that the larger the perceived risk, the lower the chosen speed will be.
Cownie and Calderwood (1966) proposed that drivers drive in a way that will main-
tain a desired level of anxiety, leading to the self-regulation of accidents within a
closed-loop; feedback from the consequences of driver decisions will affect future
decisions.
In a like manner, Wilde's risk homeostasis theory holds that drivers have a target
level of risk per unit time that they attempt to maintain. He proposed that drivers
make adjustments that ensure perceived subjective risk is equal to an internalised
target level of risk. The theory asserts that if a driver is provided with additional
safety measures, such as information concerning traffic ahead or the installation of
a seat belt, the driver will exhibit more risky behaviour to compensate and return
to the target level of risk.
Wilde also posited what he named the 'principle of preservation of the accident
rate'. This principle implies that the number of accidents within a given population
is dependent solely on the number of accidents that population is willing to tolerate.
 1. Modelling Driver Behaviour in ED and International Projects 13

1.3.4 Risk Compensation

As with other risk compensation theories, Naatanen and Summala propose a
zero-risk hypothesis, stating that drivers normally avoid behaviour that elicits
fear or anticipation of fear. Avoidance behaviour is motivated by subjective risk,
which according to the theorists is not high enough, thereby leading to accidents.
The main addition to the theory is its focus on the drivers' desired action. The
theory contends that driver behaviour is motivated not only by perception, ex-
pectancy and subjective risk, but also by the relative attractiveness and bene-
fits of carrying out a behaviour in a given situation. Furthermore, Naatanen and
Summala (1973) postulate that the motivation of desired action is the most im-
portant route leading to a driver's decision to take action. The model proposes
that a decision-making process occurs, weighing up the motivating and inhibit-
ing factors, before a decision is made for action, such as overtaking. According
to the model, driver adaptation occurs when perceptions concerning motivations
and subjective risk are altered, hence altering the balance of the decision-making
process.

1.3.5 Threat Avoidance

Fuller's (1984) threat-avoidance theory is developed both from Wilde's theory
of risk homeostasis and the zero-risk model of driver behaviour proposed by
Naatanen and Summala. The theory presupposes that drivers opt for zero risk
of accident and that they make avoidance, competing or delayed avoidance re-
sponses depending on a wide number of factors. These factors are the rewards and
punishments associated with the response, the accuracy of discriminative stimuli
recognition, the subjective probability of a threat, the effectiveness of avoidance
responses and finally the driver's level of arousal. The theory differs from the
other theories in that it is not based on a motivation variable. Instead it views
the driving task as involving learned avoidance responses to potentially aversive
stimuli.
Fuller argues against the presupposition that drivers are capable of monitoring
the probability of an accident. Instead, he proposes that drivers consider the subjec-
tive probability (or likelihood) of an accident. He proposes that the discriminative
stimulus for a potential aversive stimulus is a projection into the future formulated
through the integration of drivers' perceptions of their speed, the road environ-
ment of the intended path and their ongoing capability. Their expectations of the
threat posed by each of these factors combine, leading to either a discriminative
stimulus or no discriminative stimulus. This model also considers the influence
of rewards and punishments. He proposes that when a discriminative stimulus is
detected, the anticipatory avoidance response is not only determined by the sub-
jective probability of expected threat, but also by the rewards and punishments
associated with the various response alternatives. The theory also differs from
that of Wilde, and Naatanen and Summala in that it highlights the role of learned
responses.
14 Panou, Bekiaris and Papakostopoulos

1.3.6 Utility Maximisation

The utility maximisation model proposed by O'Neill (1977) assumes that the
driver has certain stable goals and makes decisions to maximise the expected
value of these goals. Some of these goals are achievable more effectively through
risk-taking behaviour, for example, speeding to save time or gain social status.
These motivating factors are counteracted by the desire to avoid accidents as
well as by fear of other penalties such as speeding tickets. Balancing goals with
the desire to avoid accidents therefore derives driving behaviour choice. O'Neill
claims that the balance, which affects the decision made, is shifted when a safety
measure is introduced. An assumption made by the theory, which has been ques-
tioned (OECD, 1990), is that the driver is 'rational'. In other words, the driver
is an accurate judge of the accident probability resulting from each mode of
behaviour.
Blomquist (1986) also presented a utility maximisation model, which claimed
to illustrate that risk compensation is a natural part of human behaviour when in-
dividuals pursue multiple goals with limited resources. He claimed that drivers
choose target levels of accident risk, based on the perceived net benefits of
safety effort. Again the model proposes that, under plausible conditions, a change
in safety, which is beyond driver control, causes a compensatory change in
driver effort in the opposite direction. Blomquist (1986) likens this theory to
Wilde's theory of risk homeostasis in that utility maximisation focuses on the
choice of safety goals and risk homeostasis focuses on maintenance of those
goals.

1.3.7 Behavioural Adaptation Formula

Evans (1985) proposes a human behaviour feedback parameter by which the
actual safety change in traffic systems is related to that which was expected.
A mathematical representation of the process dictates that feedback can occur
through physical changes to the system, adjustments in user behaviour for per-
sonal benefits and adjustments in behaviour to re-establish previous risk levels.
Evans suggests that human behaviour feedback is a pervasive phenomenon in
traffic systems, which may greatly influence the outcome of safety measures to
the extent that in some cases the opposite effect to that which was intended may
occur.
In a review of driving behaviour theories, Michon (1985) commented that al-
though behavioural adaptation theories explain the behaviour, and the motivations,
attitudes and factors affecting that behaviour are detailed, the actual processes
by which the behaviour occurs are not explained. Similarly, an OECD report
found the models to be vague. The report states that risk compensation theo-
ries do not explain how or which cognitions lead to the expected compensation
of objective risk. The report also points out the problem of how the objective
risk can be accurately assessed by the driver; again models fail to explain this
process.
 1. Modelling Driver Behaviour in ED and International Projects 15

1.4 Use of Driver Behaviour Models in EU and International

Projects

The field of application of such models in ED projects is vast and is gaining pace
over the last decades. Rather than attempting to meticulously cover this extensive
area, we will provide an in-depth overview of characteristic examples, showing
the related difficulties as well as benefits in applying such models.

1.4.1 Driver Models Use for Driver Training and Assessment

Indeed, many of the driver models have been developed, not aiming at driver
behaviour and support but at facilitating better, theory-based, driver training model.
A characteristic example is the GADGET-Matrix developed with the EC project
GADGET in order to structure post-license driver education.
This model as well as different extension of the Michon model have been used
as basis in nearly all recent EC projects dealing with driver training. In TRAINER
project (GRDI-1999-10024), the different layers of the GADGET-Matrix have
been further detailed and correlated to the problems of novice drivers, following
a relevant accident analysis end experts opinion survey. This work resulted in the
adapted GADGET-Matrix that correlates key subtasks of the different GADGET
layers to the needs of driver trainees, with support by multimedia tools and/or
driver simulators (Table 1.3). Thus, the development of appropriate new training
tools and scenario was based upon the relevant theoretical model of novice drivers'
needs.
Another training (in fact re-training) and assessment application is that of AG-
ILE project (QLRT-2001-00118), regarding the assessment of driving ability and
eventual re-training of elderly drivers. In this project, a mapping has been at-
tempted of the age-related deficits and benefits to the different levels and cells
of the GADGET-Matrix (Breker S. et al., 2003). The second level of the adapted
GADGET-Martix is presented in Table 1.4.
This resulted in the prioritisation of specific driving scenaria, where elderly
drivers would need more thorough assessment and/or support (re-training or
aiding).

1.4.2 Evaluation ofDriver Models' Use for Safety Aids

1.4.2.1 Use of Seat Belts
Evans (1982) observed that unbelted drivers drive at higher speeds and with smaller
headways in comparison to drivers wearing belts. This evidence supports the theory
that an adaptation in behaviour has occurred; however, the reduction in risk-taking
behaviour does not support the risk compensation hypothesis. In contrast, Streffand
Geller (1988) found that go-kart drivers wearing seat belts drove faster than non-
wearers, suggesting that the seat belt leads to a sense of security, enabling drivers
to feel safe in increasing vehicle speed. The experimental validity is questionable
16 Panou, Bekiaris and Papakostopoulos

TABLE 1.3. Findings of the analyses made in TRAINER D2.1.

Experts' proposals
Literature Existing
survey training Multimedia Simulator

Starting 0 X
Shifting gears 0 X
Accelerating/decelerating 0 X
Steering/lane following 0 X
Speed control 0 X
Braking/stopping 0 X
Use of new cars control aids (ABS, ACC, etc) X X X
Insufficient skills and incomplete automation X
Realistic self-evaluation X
Following X 0 X
Overtaking/Passing X 0 X X
Entering and leaving the traffic X 0 X
Tailgating X X
Lane changing X 0 X X
Scanning the road (eye cues) X 0 X X
Reacting to other vehicles X OX X
Reacting to pedestrians X X X
Parking 0
Negotiating intersections X 0 X X
Negotiating hills/slopes X 0 X
Negotiating curves X 0 X X
Road surface (skid, obstacles) X OX X
Approach/exit of motorways X 0 X
Turning off/over 0
Railroad crossings, bridges, tunnels 0
Reacting to traffic signs and traffic lights X 0 X X
Reacting to direction signs (including in-car X
devices)
Emergency brake OX
Urban driving 0 X X
Rural driving 0
Convoy driving X
Motorway driving OX X
Weather conditions X OX X
Night driving X OX X
Insufficient skills and incompletely X X X
automation
Information overload X X X
Insufficient anticipating skills and wrong X X X
expectations
Risky driving style X X X
Realistic self-evaluation X X X
Awareness of personal driving style X X X
Determination of trip goals, route and modal
choice
Preparation and technical check 0 X
Safety issues X 0 X
Maintenance tasks 0 X
 1. Modelling Driver Behaviour in ED and International Projects 17

TABLE 1.3. (Continued)

Experts' proposals
Literature Existing
survey training Multimedia Simulator

Intemationallegislation X
First aid (}X X
Economic driving x o X
Driver's condition (stress, mood, fatigue) X X X
Motives for driving X
Awareness of personal planning skills
Awareness of typical driving goals and risky X
driving motives
Knowing about the general relations X
between lifestyle/age/gender and
driving style
Knowing the influence of personal values X
and social background
Knowing about the influence of passengers X X
High level of sensation seeking X
Consequences of social pressure, use of X X X
alcohol and drugs
Awareness of own personal tendencies X
(risky habits, safety-negative motives)

0: the task is trained in all or nearly all European countries as the analysis of the questionnaires
showed; 0: the task is trained only in few or at least one country; X: the driving authorities and
driving instructors questioned indicate that the task is not trained, but should be trained in the particular
country.

because of the generalisation of go-karting to real life driving, as well as due to the
differences in overall behavioural patterns of drivers using seat belts versus those
that do not.

1.4.2.2 Use of Motorcycle Helmet

The retraction of these laws in some states of the USA during the late 1970s made
it possible to compare repeal and non-repeal conditions in a natural environment.
However, results from accident studies suggested that wearing helmets provided
a safety benefit; those states that had revoked laws requiring helmets to be worn
suffered an increase in fatalities (Grayson, 1996). These findings were opposed by
the analysis of Adams (1983), where higher fatality rates were found in those states
that had retained helmet-wearing laws. Adams argued that these results support the
risk compensation theory: motorcycle riders who wore helmets felt less vulnerable
to injury, thereby exhibiting riskier driving behaviours. The findings of Adams in
favour of the risk compensation theory have again been disputed by the work of
Chenier and Evans (1987). In their re-analysis of the USA accident statistics, they
found that fatalities were increased in the states that had retracted compulsory
crash helmet wearing laws. Their results imply that any increase of caution due to
TABLE 1.4. GADGET-Matrix level 2 (Mastery of Traffic Situations): Situation of older drivers (Breker et aI., 2003).
Knowledge and skills Risk factors Self-assessment

Pro Contra Pro Contra Pro Contra

Knowledge about traffic regulations/traffic Risks associated with wrong expectations/vulnerable road Awareness of strong and weak points of
signs/anticipation/communication/safety users/violations/information overload/unusual conditions/inexperience, manoeuvring skills/subjective risk
margins, etc. etc. level/subjective safety margins, etc.
-priority (esp. right before -slower on motorways - interpreting movement of other drivers -larger safety -over estimation
left, right-hand driving) -uniform driving style on -judging other drivers movement margins (headway of one's own
-signalling country roads -junctions/intersections etc.) driving abilities
- reduced perception of -larger gaps, especially - yield right of way -lower risk level -less sensibility
road signs when turning left -right-angle (side) collisions -avoiding complex to changes in
-dementia -less overtaking -left turns (right-hand driving) settings performance
-cataract -less night driving -right turns (left-hand driving) -avoiding give-way and capacities
-diabetes + associated -slower approach of - multiple vehicle accidents or stop junctions - insufficient
-glaucoma junctions -merging onto motorways -avoiding heavy awareness that
-cardiac & cardiovascular -early speed reduction at -risky merging in traffic flow traffic oneself is
condition junctions -running over red -awareness of subject to
-seizure disorders -smooth slow down at -railway crossings difficulties at general
-back pain junctions -too early observation of situations (it might changed) intersections age-related skill
-less driving on low -situation complexity especially in urban settings - general awareness declines
friction -late detection of other road users of age-related skill -dementia
-early observation of -judgement of gaps declines
situations - underestimation of speed of vehicles at higher speeds
- tolerance towards other - problems with reappearing situations when new
road users information becomes available
-problems with interrupting actions when necessary
-problems with situation complexity
-skill decline by compensation
-lateral safety margins
-dementia
-cataract
-diabetes + associated
-glaucoma
-cardiac & cardiovascular condition
- seizure disorders
-back pain
 1. Modelling Driver Behaviour in ED and International Projects 19

the removal of helmets was not great enough to compensate for the loss of safety
benefits that the helmets provide. Overall, research into motorcycle helmet wearing
has been found to provide a safety benefit and offers little support to any of the
behavioural adaptation theories. However, as Grayson (1996) pointed out these
findings are not surprising and more relevant to the theories are those mechanisms
which protect one part of the anatomy and lead to disregard for safety of other
parts of the body.

1.4.2.3 Studded Tyres

Studded tyres have been developed to improve safety in icy and snowy conditions;
they provide better track-holding properties and shorter breaking distances under
these conditions. Evidence from studies investigating the behavioural effects of
fitting vehicles with studded tyres have often been cited in support of both be-
havioural adaptation (OECD, 1990) and risk compensation (Adams, 1985). The
most greatly cited study is that of Rumar (1976). Results from this study indicated
that in icy (low friction) conditions drivers of vehicles equipped with studded tires
drive at faster speeds when negotiating curves in the road. However, it was deter-
mined that this increase in speed does not lead to a reduction in safety. It should
therefore be noted that the observation of higher speeds in itself is inconclusive.
Lund and O'Neill (1986) argue that the increased feedback provided by studded
tyres allows vehicles to be driven at higher speeds without reducing safety. The
Rumar study therefore provides evidence that the behavioural effect of driving
faster with studded tyres reduces the safety benefits; however, an overall increase
in safety is still achieved through their implementation.

1.4.2.4 Antilock Braking Systems

Antilock braking systems (ABS) are a recent safety feature introduced to the ve-
hicle, and studies concerning their effect on driver behaviour are cited to support
some of the theories of behavioural adaptation. ABS are designed to make breaking
distances shorter and to allow vehicles to be steered during breaking manoeuvres.
Rompe (1987) conducted a series of tests to investigate the benefit of these sys-
tems. Results showed that when simulating high-risk manoeuvres drivers without
ABS made 2.4 times more errors. Real life evidence does not support this size-
able predicted benefit and therefore support theories of behavioural adaptation. A
study conducted by Aschenbrenner (1994) is one of the only studies designed to
specifically investigate the risk compensation theory; their hypothesis being that
ABS will fail to reduce accidents despite its technical benefits. The study looked
at two fleets of taxis in Munich: one fitted with ABS and the control one without.
Aschenbrenner concluded that since it was not possible to prove a universal in-
crease in safety, the results indicated the occurrence of behavioural adaptation in
the form of risk compensation. The overall accident rate was unchanged, but incon-
sistencies such as decreases in blameworthy accidents and increases in parking and
reversing accidents for fitted vehicles were observed. These inconsistencies were
20 Panou, Bekiaris and Papakostopoulos

described as indicating a tendency to riskier driving by drivers of fitted vehicles

when reviewed in the GECD report.
More supporting evidence has been provided by a US study (HLDI, 1994).
Findings from this study indicated that the introduction of ABS has failed to reduce
the frequency or cost of insurance claims. However, this does not necessarily imply
that ABS does not provide a safety benefit, as Grayson (1996) points out, the
circumstances under which ABS could prevent accidents are quite rare. Kullgren
(1994) supplied evidence that ABS is effective in reducing accidents. This analysis
of Swedish accident statistics indicated an overall effectiveness of 15% for ABS
vehicles under snowy or icy conditions. It was also found that fitted cars were more
likely to be struck from behind in rear-end accidents.
Evidence for the occurrence of behavioural adaptation is both inconsistent and
conflicting. The GECD report reviewed empirical evidence concerning behavioural
adaptation. It was concluded that behavioural adaptation does occur although not
consistently, and the magnitude and direction of its effects on safety cannot be
precisely stated. The studies reviewed suggested that behavioural adaptation does
not eliminate safety gains from programmes but tends to reduce the size of the
expected benefits.

1.4.3 Driver Models Use for ADAS Design and

Impact Assessment
Few studies have been carried out investigating behavioural effects of future au-
tomated systems, and many of these have revealed the occurrence of negative
behavioural effects.
A study conducted by Winsum et al. (1989) provides support to the theory that
drivers will exhibit behavioural adaptation in response to ADAS. Winsum et al.
suggested that the use of a navigation system, in place of a map, leads to a reduction
in workload, which in tum leads to drivers increasing vehicle speed, implying that
drivers demonstrate behavioural adaptation in response to the implementation of
automated navigation.
Similarly, Forward (1993) reviewed the evidence concerning the effects of dy-
namic route guidance (DRG) systems, concluding that benefits such as reduced
workload and stress exist as do undesirable effects such as increased speed. How-
ever, Forward noted that the real effect of the system cannot be fully comprehended
until its use becomes more extensive.
Focus mainly tends to concentrate on the negative effects of introducing au-
tomation to the vehicle. However automation benefits have also been observed
during evaluation of future systems.
In the case of AWAKE (IST-2000-28062), the DRIVABILITY model has been
employed to design its warning levels and strategy. AWAKE was a project aimed
to develop an unobtrusive and personalized, real-time driver monitoring device,
able to reliably predict driver hypovigilance and effectively and timely warn
the driver. AWAKE has recognised the importance of actual traffic risk level as
well as driver status, type and key environmental factors, and worked towards a
 1. Modelling Driver Behaviour in ED and International Projects 21

TABLE 1.5. Correlation of AWAKE (driver vigilance monitoring and warning system) use
cases and warning strategy with overall DRIVABILITY Index (Bekiaris et aI., 2003).
Values of Overall
DRIVABILITY DRIVABILITY AWAKE warning
AWAKE use cases Indexes Index levels/strategy

Driver is hypovigilant; IRI = 3.5 4 (the system closely No action but

Rural environment, with WI=2 monitors the driver, monitoring system
sufficient traffic density without any action) parameters are
(normal workload); strengthened to
No major environmental EFI= 3 cautionary case
risk identified;
Standard type of driver; KSI = 2
No sign that the driver RAI=3
missed any risk.
Thus DI = 3.5
Driver is hypovigilant; IRI = 3.5 3 (the system provides Driver warning by audio
Urban environment, with WI=2 warning) and visual means
normal traffic density (warning level 1)
(normal workload);
No major environmental EFI= 3
risk identified;
Standard type of driver; KSI= 2
Driver seems to miss RAI= 2
some risks (i.e. rather
small TTC or headway).
Thus DI = 2.9
Driver is hypovigilant; IRI = 3.5 2 (the system intervenes) Driver warning by audio
Highway environment, WI=l,4 and haptic means
with low traffic density (warning level 2)
(low workload); (intervention is
High speed, cautionary EFI = 2 excluded from
case; AWAKE, due to
Standard type of driver; KSI = 2 liability issues)
Driver seems to miss RAI= 2
some risks (i.e. lane
deviation or swerving).
Thus DI = 1.75

multi-stage driver monitoring and driver warning system that takes such param-
eters into account. Table 1.5 correlates the overall DI and the indexes of the
DRIVABILITY contributors to the different AWAKE driver warning levels and
media. It should be noticed that the sensors included in the AWAKEsystem (such as
driver eyelid and steering grip force monitoring, frontal radar, lane recognition sys-
tem, etc.) allow for sufficient, real-time estimation of all DRIVABILITY indexes
(except KSI, which is however included in the system by the driver at its initiation
as the system adapts itself to the driver profile). This is done by storing driver's,
vehicle and environmental data on the system and automatically processing them.
Further processing off-line is also feasible.
22 Panou, Bekiaris and Papakostopoulos

TABLE 1.6. Driver model and rules for implementation.

PIPE drivermodel Typeof process Rules or governing assumptions
Perceptionof signals Sensorialprocess - Haptic
- Visual
-Aural
Interpretation Cognitiveprocess - SimilarityMatching
- FrequencyGambling
Planning Cognitiveprocess - Inference/reasoning
Execution Behavioural process - Performance of selectedactions/iterations

This is indeed one of the very few cases where such a direct relation between
a driver behaviour model and the development of the HMI of an ADAS has been
attempted, and in fact with great success, as the final AWAKE HMI has been rated
as adequate and useful by over 90% of its users.
Finally, within AIDE (IST-1-507674), a new driver model is being developed, at-
tempting to model concurrently the driver, the vehicle and the environment (Panou
et al., 2005). The basic assumption made for the development of the model of the
driver is that the driver is essentially performing a set of actions that are familiar ac-
cording to his/her experience. As the driving process is very dynamic, these actions
are continuously selected from a vast repository of knowledge (knowledge base) by
a diagnostic process. Consequently, the processes of diagnosis and interpretation
of acquired information become crucial for the dynamic sequencing of driver's
activity. The model of the driver adopted is based on a very simple approach that
assumes that behaviour derives from a cyclical sequence of four cognitive func-
tions: perception, interpretation, planning and execution (PIPE). This model is
not sequential as the execution function, i.e., the manifested form of behaviour,
may result from several iterations (cyclical) of the other functions. Moreover, in
agreement with the initial hypothesis, the planning function, is usually bypassed
by the 'automatic' selection of familiar frames of knowledge that are associated
with procedures or sets of several actions aiming at the fulfilment of the goal of a
frame. This function is important as it becomes effective in unknown situations or
in the case of novice drivers, when 'simpler' frames, based on single actions or on
a limited sequence of very simple/familiar actions, are called into play to deal with
the situation. These four cognitive functions can be associated to either sensorial
or cognitive processes and are activated according to certain rules or conditions
(Table 1.6). This model will be utilised in personalising the multi-ADAS system
HMI, in accordance to a particular driver's needs and preferences, and will also
be used in the traffic environment.

1.5 Conclusions
Driving task modelling has started as simple task-layers representation for taxo-
nomic use in driver training and has gradually evolved to dynamic models, which
consider driver behaviour adaptation as well as the impact of the traffic environment
 1. Modelling Driver Behaviour in ED and International Projects 23

and the driving context. The initial list of driver training and assessment projects
that used driver models as their theoretical basis (GADGET, DAN, TRAINER,
AGILE, CONSENSUS, etc.) have been followed by a new generation of projects
that use driver models to assess the impacts of driver support systems (i.e.,
ADVISORS, TRAVELGUIDE) and, more recently, by those that attempt to use the
model parameters for optimal HMI design (AWAKE, COMUNICAR, AIDE, etc.).
Preliminary results have proven that such a correlation is feasible and beneficiary,
but it is far from obvious. The model output has to be evaluated and even modi-
fied by empirical results. Thus, currently the model is being applied and tested in
AIDE, SENSATION and PREVENT Integrated Projects through short- and long-
term testing of drivers. Furthermore, the model can only be at the starting basis
of the design and development process and only influence the actual ADAS HMI
within predefined design boundaries. Nevertheless, we seem to be at the infancy
of a new design principle for driver support training and assessment systems - the
model-based modular and personalised design.

References
Adams, 1.G.U. (1985). Smeed's Law, seat belts and the emperor's new clothes. In
L. Evans and R.C. Schwing (Eds.). Human Behaviour and Traffic Safety. Plenum,
New York.
Aschenbrenner, K.M. and Biehl, B. (1994). Empirical studies regarding risk compensation
processes in relation to anti-lock braking systems. In R.M. Trimpop and GJ.S. Wilde
(Eds.). Challenges to Accident Prevention: The Issue of Risk Compensation Behaviour.
Styx Publ., Groningen, The Netherlands.
Bainbridge, L. (1987). Ironies of automation. In 1. Rasmussen, K. Duncan and 1. Leplat
(Eds.). New Technology and Human Error. Wiley, New York.
Baron, S., Zacharias, G., Muralidharan, R. and Lancraft , R. (1980). PROCRU: A Model
for Analyzing Flight Crew Procedures in Approach to Landing. NASA CR-152397.
Bekiaris, E., Papakonstantinou, C., Stevens, A., Parkes, A., Boverie, S., Nilsson, L.,
Brookhuis, K., Van Wees, K., Wiethoff, M., Damiani, S., Lilli, F., Ernst, A., Heino, A.,
Widlroither, H. and Heinrich, J. (2001). ADVISORS Deliverable 3/8.1v4: Compendium
of Existing Insurance Schemes and Laws, Risk Analysis ofADA Systems and Expected
Driver Behavioural Changes. User Awareness Enhancement, dissemination report and
market analysis and ADAS marketing strategy. ADVISORS Consortium, Athens, Greece.
Bekiaris, E., Amditis, A. and Panou, M. (2003). DRIVABILITY: A new concept for mod-
elling driving performance. International Journal ofCognition Technology & Work, 5(2),
152-161.
Blomquist, G. (1986). A utility maximization model of driver traffic safety behaviour.
Accident Analysis and Prevention, 18,371-375.
Boverie, S., Bolovinou, A., Polychronopoulos, A., Amditis, A., Bellet, T., Tattegrain-Veste,
H., Manzano, 1., Bekiaris, E., Panou, M., Portouli, E., Kutila, M., Markkula, G. and
Angvall A. (2005). AIDE Deliverable 3.3.1: AIDE DVE monitoring module - Design
and development. AIDE Consortium, Toulouse, France.
Breker, S., Henrikson, P., Falkmer, T., Bekiaris, E., Panou, M., Eeckhout, G., Siren, A.,
Hakamies-Blomqvist, L., Middleton, H. and Leue E. (2003). AGILE deliverable 1.1:
Problems of elderly in relation to the driving task and relevant critical scenarios.
24 Panou, Bekiaris and Papakostopoulos

Cacciabue, P.C., Mauri, C. and Owen, D. (1993). Development of a model and simulation
of aviation maintenance technician task performance, International Journal ofCognition
Technology & Work, 5(4),229-247.
Chenier, T.C. and Evans, L. (1987). Motorcyclist fatalities and the repeal of mandatory
helmet wearing laws. Accident Analysis and Prevention, 19, 133-139.
Christ et al. (2000). GADGET final report: Investigations on influences upon driver be-
haviour - Safety approaches in comparison and combination. GADGET Consortium,
Wien, Austria.
Cownie, A.R. and Calderwood, 1M. (1966). Feedback in accident control. Operational
Research Quarterly, 17, 253-262.
De Vos, A.P., Hoekstra, W. and Hogema, I.H. (1997). Acceptance of automated vehicle
guidance (AVG): System reliability and exit manoeuvres. Mobility for everyone. Pre-
sented at the 4th World Congress on Intelligent Transport Systems, Berlin.
Evans, L. (1985). Human behaviour feedback and traffic safety. Human Factors, 27 (5),
555-576.
Evans, L. and Schwing, R.C. (Eds.) (1982). Human Behaviour and Traffic Safety. Plenum,
New York.
Forward, S.E. (1993). Prospective methods applied to dynamic route guidance. In O.MJ.
Carsten (Ed.). Framework for Prospective Traffic Safety Analysis. HOPES Project De-
liverable 6.
Fuller, R. (1984). A conceptualization of driving behaviour as threat avoidance. Ergonomics
27, 1139-1155.
Grayson, G. (1996). Behavioural adaptation: A review of the literature. TRL Report 254.
Transport Research Laboratory, Crowthorne.
Hatakka, M., Keskinen, E., Gregersen, N.P., Glad, A and Hernetkoski, K. (1999). Results
of EU-project GADGET. In S. Siegrist (Ed.). Driver Training, Testing and Licensing-
Towards Theory Based Management ofYoung Drivers' Injury Risk in Road Traffic. BFU-
report 40, Berne.
HLDI. (1994). Collision and property damage liability losses ofpassenger cars with and
without antilock brakes. Highway Loss Data Institute Report A-41. HLDI, Arlington,
VA.
Howarth, I. (1993). Effective design: Ensuring human factors in design procedures. In
A. Parkes and S. Franzen (Eds.). Driving Future Vehicles. Taylor and Francis, London.
Kullgren, A., Lie, A. and Tingvall, C. (1994). The effectiveness of ABS in real life accidents.
Paper presented at the Fourteenth International Technical Conference on the Enhanced
Safety of Vehicles, Munich.
Lund, A.K. and O'Neill, B. (1986). Perceived risks and drinking behaviour. Accident Anal-
ysis and Prevention, 18, 367-370.
McKnight, AJ. and Adams B.D. (1970). Driver Education Task Analysis, Vol. 1: Task
Descriptions. Human Resources Research Organization, Alexandria, VA.
Michon, lA. (1985). A critical view of driver behaviour models: What do we know, what
should we do? In L. Evans and R.C. Schwing (Eds.). Human Behaviour and Traffic Safety.
Plenum, New York.
Naatanen, R. and Summala, H. (1973). A model for the role of motivational factors in
drivers' decision making. Accident Analysis and Prevention, 6, 243-261.
OECD. (1990). Behavioural Adaptations to Changes in the Road Transport System. OECD,
Paris.
O'Neill, B. (1977). A decision-theory model of danger compensation. Accident Analysis
and Prevention, 9, 157-165.
 1. Modelling Driver Behaviour in EU and International Projects 25

Panou, M., Cacciabue, N., Cacciabue, P.C. and Bekiaris, E. (2005). From driver modelling
to human machine interface personalisation. Paper presented at the IFAC World Congress,
Prague.
Ranney, T. (1994). Models of driving behaviour: A review of their evolution. Accident
Analysis and Prevention, 26(6), 733-750.
Rasmussen, J. (1984). Information processing and human-machine interaction. In An ap-
proach to cognitive engineering. North Holland, New York.
Rompe, K., Schindler, A. and Wallrich, M. (1987). Advantages of an anti-wheel lock system
(ABS) for the average driver in difficult driving situations. Paper presented at the Eleventh
International Technical Conference on Experimental Safety Vehicles, Washington, DC.
Rumar, K., Berggrund, U., Jernberg, P. and Ytterbom, U. (1976). Driver reaction to a
technical safety measure - Studded tyres. Human Factors, 18, 443--454.
Stokes, A., Wickens, C. and Kite, K. (1990). Display Technology: Human Factors Concepts.
Society of Automotive Engineers, Inc., USA.
Streff, F.M. and Geller, E.S. (1988). An experimental test of risk compensation: Between-
subject versus within-subject analyses. Accident Analysis and Prevention, 20, 277-287.
Summala, H, Lamble, D. and Laakso, M. (1988). Driving experience and perception of
the lead cars braking when looking at in-car targets. Accident Analysis and Prevention,
30(4), 401--407.
Taylor, D.H. (1964). Drivers' galvanic skin response and the risk of accident. Ergonomics,
7, 439--451.
Van Winsum, W., Van Knippenberg, C. and Brookhuis, K. (1989). Effect of navigation
support on driver's metnal workload. In Current Issues in European Transport, Vol. I:
Guided Transport in 2040 in Europe. PTRC Education and Research Services, London.
2
TRB Workshop on Driver Models: A
Step Towards a Comprehensive Model
of Driving?
DELPHINE CODY AND TIMOTHY GORDON

2.1 Introduction
Various disciplines use the same or similar terminology for driver models - vehicle
and traffic engineering, psychology, human factors, artificial intelligence to men-
tion the most common; however, the definition of the term varies not only between
disciplines but even between different researchers within any given discipline. Re-
cent efforts in applied psychology and human factors have emphasised the need
of developing models that can be implemented and used in computer simulation,
hence representing a possible link between these disciplines, and also a chance
to consider the broader picture of driver model within a transportation/traffic sys-
tem. In order to discuss this link, the authors organised a workshop on driver
modelling during the 84th annual meeting of the Transportation Research Board,
Washington, DC. The workshop was attended by 25 researchers from the vari-
ous fields listed above and lead researchers from the United States, Europe and
Asia.
Three objectives were set for this workshop: (i) create a group of driver model
developers and users, (ii) share common experience and reach common definitions
and finally (iii) set a road map for the next generation of driver models. The topics
that were more specifically dealt with were the design and application of cognitive
and driver behaviour models and their integration within a broader simulation
framework. This chapter is divided into three sections. In the first section, we
present the workshop content in more detail and provide a summary of the speakers'
contribution. In the second section, we present a synthesis of these models with
the introduction of a set of dimensions allowing for visualisation of the different
models on a similar scale. The final section presents what the authors believe to
be the critical steps necessary to coordinate efforts towards a new generation of
driving model- a framework where researchers from different fields contribute to
create a comprehensive model of the driving activity.

26
 2. TRB Workshop on Driver Models 27

2.2 Workshop Presentation and Speakers' Contribution

This section comprises two parts. We will first detail each of the themes around
which the speakers organised their presentation and will present summaries of
each of the presentations in the second part.

2.2.1 Workshop Content

The workshop was based on three themes: (i) driver model purpose and applica-
tion, (ii) driver model architectures and implementations and (iii) driver model
calibration and validation. We will now discuss each of these themes.

2.2.1.1 Driver Model Purpose and Application

From a general standpoint, generating driver models can be seen as equivalent
to developing a comprehensive description of scientific knowledge about drivers.
The first step towards making a hierarchy of scientific methods to approach a phe-
nomenon such as the driving activity is to conduct specific studies to observe and
understand the phenomena (e.g., drivers' glances, perception of speed, decision-
making process, drivers' attention control) and the second step is to reconcile the
results of the different studies into a comprehensive description - in other words
a model dynamically linking the results of these different studies to reproduce the
driving activity. Hence, one of the interests of building driver models is to vali-
date the results of studies looking into specific aspects of driving and identify the
aspects that need to be addressed to further the understanding of driving.
Another interest for developing such models resides in their application. Two
types of applications were discussed during the workshop: (i) application to safety
and driver assistance system design and (ii) prediction and evaluation of new
systems on traffic flow. Three presentations fell under the first type, namely, 'In-
Vehicle Information System (IVIS) Model' by Jon Hankey, 'ACT-R Driver Model'
by Dario Salvucci and 'Workload in Driver Modelling' by Jeroen Hogema and
Richard van der Horst. The two other presentations, 'ACME - Driver Model' by
Daniel Krajzewicz and 'FLOWSIM' by Mark Brackstone, addressed the traffic
assessment category.

2.2.1.2 Driver Model Architecture and Implementations

Model architecture can be seen as the blueprint of the model. In that sense, a
simple description of a model (see Fig. 2.1) consists of three elements: inputs,
information processing or behaviour and outputs. This simple architecture allows to
generate parameters of factors that can be used to derive measures of effectiveness
or evaluation.
28 Codyand Gordon

/ Driver Models Information Processing

' Perception

/
Inputs Outputs

~
' Cognitive Processes
' driver assistance systems -Hurnan motor control
Vehicle

I
' other vehicles
'road geometry and signals -Decisions

Driver behavior
Vehicle control
\.
1
FactlJrs/ parameters

Measures of Effectiveness/Evaluation
Risk acceptance
safety
I--

Cras h risk
Efficiency
Workl oa d
Stress

FIGUR E 2.1. Modularhigh level representation of drivermodel structure.

2.2.1.3 Calibration and Validation

Model calibration consists of setting values for parameters which support the sim-
ulation. For example, when a model includes a gap regulation task, parameters that
need to be calibrated are the gaps that the simulated drivers will maintain as well
as conditions for variation of these gaps. The calibration addresses elements of
the driving tasks as well as cognitive processes (e.g., memory decay) or other ele-
ments of the simulation (e.g., vehicle models) . We distinguished two main methods
for calibrating model s. The first method consists of using data and results avail-
able in the scientific community, and the second consists of running specific data
collection. This method can be subdivided into three more categories, depending
on whether the data collected is driver centric (gaze, attention) or vehicle cen-
tric (steering, braking, speed contro l) or traffic centric (lane keeping , range , range
rate) . The use of the first method brings up the issue of standardisation in method
to gather and measure data and the definition of parameters that are derived .
For model validation, there is one method of comparing the model output with
data sets. The variation in the application of the method depends on the type of
data used. For instance, the model output can be compared with collected data,
with on-going behaviour, or with more general data . The comparison of model
output with on-going behaviour raised an interesting question on the nature of
the model outp ut. Is a model providing a trend or should we expect to be able
to predict real-time behaviour? The question was not answered directly during
the workshop. Although some researchers presen t in the audience thought that a
prediction of driver behaviour in real time would be a huge achievement for driver
models, a lot of doubt about the possibility to reach such a level of accuracy was
also expressed.
Clearly, there is fundamental issue resulting from the inherently stocha stic na-
ture of driving . On the one hand, a ' real-time' model should be capable of making
precise and detailed predictions of driver actions as a function of time. On the other
hand, real drivers display a range of decisions, driving styles, levels of attention,
 2. TRB Workshop on Driver Models 29

etc., and so should any realistic driving model; resolving these apparently conflict-
ing viewpoints requires a somewhat deeper concept of what is required for model
validation. At the very least, validation of real-time models must take account for
a wide range of stochastic influences.
Out of the five models that were presented, four are proceeding with more or
less extensive validation and one is still in the process of calibration.

2.2.2 Summaries of the Speakers' Contributions

Five speakers presented their research relative to each of the themes: Jon Hankey,
Virginia Tech Transportation Institute; Dario Salvucci, Drexel University; Daniel
Krajzewicz, Deutsches Zentrum fur Luft - und Raumfahrt e.V. (DLR); Mark Brack-
stone, School of Engineering and the Environment, University of Southampton,
and Richard van der Horst, TNO Human Factors. All the five summaries are di-
vided into three parts: (i) model purpose description, (ii) architecture description
and (iii) calibration and validation issues. The architectures that were presented are
briefly explained for each of the model. The reader may refer speakers' publica-
tions on their model for more thorough descriptions. The models are also discussed
based on the following factors: (i) simulation scale, along two dimensions, i.e., the
unit simulated (e.g., one driver, groups of drivers or vehicle fleets) and the number
of unit that can be simulated at once; (ii) visualisation, in terms of drivers' states,
parameters that can be observed in simulation and traffic; (iii) the ease in using
these models by other researchers.

2.2.2.1 In-Vehicle Information System - Jon Hankey

The purpose of in-vehicle information system (IVIS) is to provide a proof of
concept design tool in order to assess and compare in-vehicle system concepts,
prototypes and products for system designers and to provide insight into potential
design improvements. The approach for the development of this model consists of
compiling, analysing and using data from actual empirical research for creating
'micro-models' of performance prediction. The model predicts the visual and
cognitive resources needed at a basic level. These predictions are made at the
subtask level, because subtasks often require different resources and subtasks can
be combined into tasks which can themselves be combined into system assessment.
Therefore, the model accounts for a variety of information input, information
processing and response output combinations.
The overview of the IVIS demand model is shown in Fig. 2.2. During the
workshop, the emphasis was put on the software development of the tool and how
users would manipulate the tool. In terms of driver model, if we observe Fig. 2.2,
the component that most closely relates to the information processing is the one
listing driver resources involved, i.e., visual demand, auditory demand, manual
demand and speech demand. The assumption of the model is that the driver is a
finite capacity, with a single-channel processor of visual information. The driver
can 'share' cognitive resources to differing degrees and the resources needed to
30 Cody and Gordon

Task Specification in Terms of Driver Resources Involved

Visual Demand Manual Demand
Auditory Demand Speech Demand
Supplemental Information Processing Demand
I

Comparison/Interpolation Recall Task From

to Task(s) Found in Library Previously Saved Analysis

Program Identifies Task-Relevant Nominal Values for Primary Measures

Modification of Values, As Desired

Task Specific Subtask Specific

TaskllVIS Evaluation Using Figures of Demand

and Derived Values for Primary Measures

FIGURE 2.2. Overview of IVIS demand model.

perform IVIS tasks compete with the resources needed to drive the vehicle. In order
to avoid reducing the driving performance to unacceptable levels, only limited
resources can be required by IVIS. The different resource components and/or
magnitudes can be used to perform the same task and are required for different
tasks. The combination of the demand of these specific components determines the
required resources to perform a task. The required resource demand can be used to
estimate the potential of a decrement in driving performance. The principle of IVIS
is that nominal values for measures are derived that can be modified to match a
task or design specification; for example, a subtask modifier is the message length
and a task modifier is the roadway complexity.
The scale of simulation is one driver within one of the three age groups. The
software allows visualising the behaviour change outcome and the modifiers that
can be selected to adjust. Finally, the source code and a user manual are available
from FHWA. The software is still in a proof of concept stage. The model is used and
designed to be used in the industry, which can make the feedback to the scientific
community a slower process. For a review of the validation carried out at VTTI,
the reader can consult Jackson and Bhise (2002).

2.2.2.2 ACT-R Driver Model- Dario Salvucci

The theoretical endeavour of the ACT-R driver model is to provide a psycholog-
ically plausible model of an individual driver and combine cognitive, perceptual
 2. TRB Workshop on Driver Models 31

Declarative Module
(TemporallHippocampus

External World
FIGURE 2.3. ACT-R architecture.

and motor dimensions (Salvucci et al., 2001) . The approach involves applying
a cognitive architecture, ACT-R associated to a computational framework . The
advantage of this approach is the possibility to re-use theories and mechanisms
already integrated within the cognitive architecture. This approach currently fo-
cuses on highway driving involving moderate traffic. The two current practical
applications are (i) the prediction of driver distraction, for which models of typical
secondary tasks (e.g., phone dialling) are integrated into the model in order to pre-
dict real-world observables measures and (ii) the recognition of driver intentions,
where many models are run simultaneously, each trying to accomplish a different
goal and the method consist of tracking which model best matches the observed
data.
Figures 2.3 and 2.4 depict two diagrams: one for ACT-R and the othe r for ACT-
R driver model. Figure 2.3 shows (Anderson et a\., 2004) ACT-R architecture
and details the overlap between brain structures and information processing steps.
Figure 2.4 describes (Salvucci et a\., 2001) the component involved in the ACT-R
driver model. ACT-R provides a framework and specific rules are developed for
the driving model. The model also integrates cognitive models for other tasks, such
as using a cell phone or a navigation system. The current scale of simulation for
the ACT-R driver model is focused on one driver in an environment that provides
32 Cody and Gordon

With probability P1lU11dtor Iflead caris too close

FIGURE 2.4. ACT-R driver model representation.

interactions with other vehicles. This effort also includes the modelling of younger
versus older drivers. The visualisation allows observing driver's eye view and
mirror and in a new system currently under development, graphs will provide
measure of behaviour. Regarding the ease in using a model, the current version
of the ACT-R driver model cannot be used easily by other researchers, but a new
version is under development that should be easy to use (Salvucci et al., 2005).
In terms of calibration and validation, the approach applied for the ACT-R driver
model is the comparison of model and human data sets recorded similarly in terms
of real-world measures. The data sets that have been collected so far are highway
driving, phone dialling and distraction, radio tuning and distraction and the age
effects. Examples of measures are lane change steering profiles, gaze distribution
on highway, lateral deviation during phone dialling.

2.2.2.3 Optimal Control Model- Richard van der Horst

The driver modelling effort at TNO aims at evaluating driver behaviour, perfor-
mance and workload and at providing inputs for traffic flow model, such as MIXIC
driver model and human-kinetic traffic flow model. This approach focuses on indi-
vidual driver and vehicle units and is based on experimental research. The frame-
work applied for this modelling effort is the one of 'optimal control', i.e., the use
of a linear system theory, where the assumption is that the driver is well trained and
well motivated to behave optimally. The model also integrates inherent limitations
and constraints. The resulting model is a realistic description of driver behaviour
in terms of driver performance, workload and total system performance measures.
 B-1
2. TRB Workshop on Driver Models 33

DRMR H VEHICLE ~ BEHAVJOR

f
FI GURE 2.5. Block diagram for driver/vehicle/system and driver/model.

Two figures support the model architecture. Figure 2.5 is a block diagram of
the driver vehicle system. In this figure , the concept of task covers elements such
as lane keeping , car following or speed control , while the behaviour is described
as measures of performance and workload . Figure 2.6 illustrates a 3 x 3 x 3
description of driver behaviour (Theeuwes, 2001) relative to the task hierarchy,
task performance and information processing . The shaded area denotes the part
of the behaviour that is currently modelled , i.e., the control of the vehicle at a
skilled-based level. For this part of the behaviour, all of the information proce ss
is integrated, from percept ion to action. Figure 2.6 provoked a lot of interest from
the audience, as it is commonly considered that strategic equal s knowled ge-based
level, manoeu vring equal s rule-b ased level and control equals skill-based level. A
question was raised about how to transition within this 3 x 3 x 3 representation.
For example , how can a driver be simulated at the skill-based control and then
simulated at the knowledge-based level? Is it only by learning? Can it also be
due to other factors? The answer pointed to situations such as degraded driving
co nditions. For exampl e, an experienced driver mobili ses skills (Task performance)
at the control level (Task hierarchy), but while driving another vehicle will move
the vehicle control one level up to the rule level. Another example given was the
case of a novice driver, for whom the control of the vehicle can be associated to a
kn owledge-level perform ance.

>-
/
:I: ,/
U ,/ / v
0:::
<
0:::
strateg ic ///

,./.
W maneuvering A /l.-(
:I:
~ control

s
In

FI GURE 2.6. Driver model.

34 Cody and Gordon

World
IObjects I 4--------------------------------------------------------
vehicle dynam ics

---------1

r
Driver ~ - - -- , , --- - -- ., ,... ---
Sensors
I
Processor I
Body
Eyes, 1 s ensor-r H Action- Anns , Legs ,
Ears. filters filter Head
- - _... , I, , -- - -
Proprioception
---- ---,
I
.a.
,
I I ..., 0

---- ----- ---- ----------,

0
- - -- - ----- - -- - ------- -- ~
attent ion control haptic feedback

Vehicle
--------- --- ----- ------ --- ---- ---- ----- ----- --- -- - Controls
Field of view Pedals ---
Steering whee l.. .

FIG URE 2.7. ACME driver model.

The outputs generated by the OCM are time simulation results , performance
and workload measures and a workload index. The example of validation provided
during the workshop was a comparison of data collected on a driving simulator and
of model prediction on the lane-keeping task . The agreement between the model
and experimental results indicates a useful predictive capability.

2.2.2.4 ACME
The purpose of the ACME driver model it to develop a combined view on car-
following and lane-changing as well as describe and evaluate the mental processes
needed for driving . It is a man in the loop simulation of a 'car-driver' unit composed
of three sub-models: model of human sensors, model for information processing
and a model for action execution. The simulation integrates models of the dynamics
of the vehicle and models of the simulated area.
The ACME driver model is a very modular architecture (Fig. 2.7) and one of the
most driver vehicle infrastructure system oriented among the presentations given.
The model can be divided into three major substructures: (i) senses, (ii) informa-
tion processing and (iii) actions execution. The senses that are implemented are
vision and hearing, and the perception of acceleration is used to influence speed
decision and the haptic input is not integrated. The information processor consists
of an internal world representation storing , a planning instance and an execution
instance. In order to carry out the execution of action, the extremities are simulated
to move in-vehicles devices to certain position. These devices in tum determine the
vehicle's dynamism. Regarding the scale of simulation, approximately 20 vehicles
can be simulated around an intersection. The simulation step ranges from 10 to
100 ms. The visualisation represents the states within the driver's cognition and
includes timelines of measures. Regarding the ease of using the model, it would
 2. TRB Workshop on Driver Models 35

While driving, you are...

Watc hing

In Fuzz y Logic, this becomes

Fuzzy Input Fuzzy Firing Defuzzification

x (relative speed) =A' If x =A and y = B then z = C Because x =A' and y = B'

y (headway divergence) = B' z (acceleration rate) = C'

FIGURE 2.8. FLOWSIM driver model.

still require time for the user to become familiar with the model and its implemen-
tation. The ACME model is still in the process of development and hence has not
been validated yet.

2.2.2.5 Fuzzy Logic Based Motorway Simulation

The development of Fuzzy LOgic based motorWay SIMulation (FLOWSIM)

started in 1997. The model was initially focused on highway traffic, but now covers
all type of roadway. This framework was first used for advanced driver assistance
system (ADAS) and infrastructure speed control applications. Recent urban appli-
cations include network travel time prediction. It is now the object of an intensive
enhancement program in China and has been chosen as the traffic simulation tool
for supporting the city of Tianjin's transportation planning and management.
FLOWSIM driver model (Brackstone, 2000; ;rackstone et aI., 1997 ; Wu et aI.,
2000) representation (Fig . 2.8) presents the three basic steps of watching, thinking
and responding and how these basic steps can be associated to fuz zy logic. The
example used for the description of the model focused on speed and gap control. It
showed how the relative speed and distance divergence influence the driver 's action
in terms of acceleration rate . This model allows most units to be simulated during
one simulation, with the possibility of simulating up to a 1000 vehicles. Different
behaviours are simulated by using distributions . The speaker found little empirical
evidence about the existence of groups such as young or old or aggressive drivers
and therefore uses the distribution without labels tying the driver/vehicle unit to
specific groups. The visualisation mainly displays individual vehicles and the road
geometries. In order to use the system, it would be necessary to receive training.
The validation of FLOWSIM (Wu et aI., 2003) concentrated on the comparison
of simulated and measured traffic flows and showed a very satisfying fit.
36 Cody and Gordon

2.3 Synthesis of Presented Models

This section presents two sub-sections. In the first one, we introduce a set of
dimension and visualisation methods to understand the scopes of the different
models that were presented above, their commonalities as well as their dif-
ferences. In the second section, we will discuss briefly the notion of driver
model as a tool used for the development or evaluation by users who are not
necessarily human sciences scientists and, although they need to manipulate
driver model, should not have to create all of the elements going into a driver
model.

2.3.1 Understanding Models' Scope

We consider that the basic architecture of a simulation involving a driver
model is constituted of at least three main components: (i) a driver model,
(ii) a vehicle model and (iii) an infrastructure model. This hierarchy (Fig. 2.9)
can be further expanded into increasing the levels of detail. For example, the
driver model can be expanded in more modules or dimensions, such as driv-
ing tasks classification (strategic, tactical and operational) or psycho-motor
dimension, which is constituted of perception, cognition and motor control.
For the workshop, we defined a number of dimensions for each of these
three main components and asked the speakers to rate their models on these
dimensions.

Driving tasks Vision

Driver model perception

Psycho-motor Auditory
cognition

Motor control
Vehicle response
Vehicle(s) model

Vehicle motions

level of details

Infrastructure model

Type

FIGURE 2.9. Hierarchy of model components.

 2. TRB Workshop on Driver Models 37

For the driver component, the proposed dimensions were:

psycho-motor
driving tasks
control level
model capacities
simulated phenomena
simulation control

For the vehicle component, the proposed dimensions were

control variables
vehicle motions
vehicle subsystems
vehicle response

And, for the infrastructure component, the two main proposed categories were

level of details (lane, signalisation)

environment type (highway, urban ....)

As the workshop was limited in time, we focused the scope of the description to
the three main components and on the details of the psycho-motor dimensions.
The rating was as follows:
0: not represented
1: indirectly represented (there is a related parameter)
2: basic inclusion (there is a specific parameter in the model that indicates relevant
trends)
3: included (the parameter or state is directly represented via a simple sub-model)
4: modelled (a sub-model represents the process)
5: represented in detail (the process is a core aspect of the model and be related in
some detail to experiments or theories)
This rating was then integrated in a 'radar web' graph. Each speaker presented
a graph for each of the three components and then a more detailed graph of the
psycho-motor level. The ratings for all the models were then integrated on a same
graph (see Figs. 2.10 and 2.11).
The purpose of representing the models presented at the workshop via a common
graph was to identify synergies and limitations. In order to categorise synergistic
dimensions, it was proposed that when a dimension was ranked three or above
by at least three models, then a synergy between the three models is possible. As
shown in Fig. 2.10, the dimensions that fit this category are perception, cognition,
operational, rule, reaction and static input. A second category is one where di-
mensions could clearly be further developed, where at least one model ranks three
or above. The dimensions in this category are motor control, strategic, tactical,
skill, non-driving activities, knowledge, anticipation, distraction and driver char-
acteristics. Another interesting category is one where none of the models ranked
 38 Codyand Gordon

Perception Ps cho-motor

Simulation
Control

Driving
driver characteristics tactical Tasks

reaction time~~~~~~~~~~~~~l;;;9~~~Joperational
Simulated

Control Levels
Model Capacities
non driving activities

IVIS ..... ACT-R ..... Workload - ACME ~ Flowsim I

FIG URE 2.10. Representation of drivermodel dimensions.

three or above with the following three dimensions: learning , variable and dynamic
inputs.
The same method was applied for the description of the psycho-motor dimen-
sions (Fig. 2.11), and the areas for which we identified synergies are vision, decision
making , memories, steering control and velocity control. The areas that need to
be developed are haptic, auditory, recognition, anticipation, rules/knowledge and
attentional resources. Finally, the areas still to be covered are tactile and other
vehicle control.
This categorization of the dimensions led to a question about researchers imple-
mentation strategy and what dictates the choice of what to implement first. Are the
dimensions from the first category receiving so much attention because they are
easier to implement than the low ranking ones or because they are more important
to the concept of driver model and its current applications?

2.3.2 Driver Model Toolbox

The notion of a driver model toolbox came up during the workshop discussion. A
parallel was drawn with tools such as Matlab and Simulink, which allow the use
of pre-existing tools to create simulations; for driver models, researchers have yet
to start developing such a simulation environment. The two main aspects covered
were how the toolbox should be developed and how to identify potential user
group s.
 2. TRB Workshop on DriverModels 39

Vision
Perception

Motor Control

steering contro' 1\~f-l-:l-~~SJ~ ~=7==;;P'i- l

attentional resources

Cognition
I...... IVIS " ' - ACT-R ..... Workload ACME .... Flowsim I

FIGURE 2.11. Representation of psycho-motor dimensions.

Matlab is a commercial software package supporting simulations from most en-

gineering disciplines, and it is difficult to aim at a completely similar product. The
equivalent would be a commercial software supporting simulation of intelligent
systems or of the human mind . One of the open suggestions was to create a pub-
licly available tool, based on the contributions of an international consortium of re-
searchers. Such a tool would be a ' driver simulator' software, which would allow to
observe as a simulation is conducted the different components of the model activity .
The two key groups - yet to be identified - are a scientist/researcher group,
who would design and develop the toolbox, and a parallel user group, who would
test, develop and help validate the resulting models . It is anticipated that whilst
these two groups would have a range of differing requirements and interests,
their activities could be complementary and lead to a new general level of driver
modelling capability.
In some respect, the models closest to achieving this role are the IVIS and
ACT-R driver model. The limitations of these models are their limited scope , on-
going calibration and validation and for the ACT-R driver model, the step of going
towards a more user friendly platform and interface.

2.4 Towards a Comprehensive Model of Driving

Any comprehensive model of driving should be capable of reproducing statisti-
cally verifiable trends in response to changes made to its constitutive parameters,
40 Cody and Gordon

and those parameters should be rich enough to encompass the interests of a very
wide class of transportation researchers - across the disciplines mentioned in the
Introduction (Section 2.1). Roughly speaking, relevant time-based responses (eye
glance, steering, emergency braking, decision latency, etc.) should be available to
predict trends as key parameters of interest change (external vehicle behaviour,
vehicle control system activity, driver experience, vehicle information and warning
systems, driver-vehicle interface, distracting activities, external signalling, etc.).
No such model exists at present; indeed, it is probably naive to expect that any
single model will ever be sufficiently full and complete to represent all possible
behavioural changes in response to all possible parametric changes. However it is
plausible to expect that a single modular formulation of the driving process - or
more precisely a single functional representation of driving - could be developed
to accelerate the progress on any particular model-based question of this type.
A common modular framework is also an essential starting point for any se-
rious coordinated effort in driver modelling. In the foregoing text (Fig. 2.9) we
have suggested the general scope of such a modular approach, but the framework
itself is undefined. The aim here is to start to define what such a framework might
look like, without necessarily proposing this in any final form. Figure 2.12 illus-
trates the point. Each rectangular shape depicts a process that is at the same time
stochastic and predictable. Those associated with the driver are open to learning
and adaptation. Note that in this representation, the driver appears as a distributed
set of processes! Indeed, the structure embodies a whole range of assumptions and
hypotheses. For example, it suggests that some form of 'vision-based' driving is
possible without the involvement of higher cognitive function, but that manoeu-
vring decisions are not possible without (at least occasional) strategic input. The
diagram is not intended to be complete (e.g., there is no auditory information chan-
nel) or even correct; indeed the correctness of this or other functional maps of its
kind ultimately depends on its ability to match experiments.

2.5 Conclusions
The models presented at the workshop varied based on the goal for which they
are designed and the methods applied to implement, calibrate and validate them.
However, they do share commonalities in the processes that they manipulate. The
authors' intention when convening the speakers was to illustrate the variety of
driver models and to convey the point that the aim of developing driver models is
not to create 'the' driver model or a driver model representing the right approach.
Actually, the level of detail to integrate in a driver model really depends on the goal
of the developer. For instance, in order to predict driver behaviour, how necessary
is it to describe the psycho-motor processes underlying the driving activity or is a
data analysis of data patterns and trends sufficient?
The intent of the workshop was to open up the discussion on modelling issues
and exchange, move from a 'researcher centric' to a community-directed approach
in order to take driver model development further. The other goal was to discuss the
 2. TRB Workshop on Driver Models 41

strategic

DVI

vehicle vehicle
control
dynamics kinematics

:=
+-- other
vision vehicles

infrastructure

FIGURE 2.12. Possible functional representation of the driving process.

idea of extending the notion of driver model towards the concept of comprehensive
driving model, where the driver becomes one the component of a system and could
become a tool used in order to support ITS development or other safety application
for which it is necessary to know more about the driver behaviour or information-
processing characteristics. In this sense, this concept differs from classical traffic
simulation tool by the scale at which it is envisaged, although simplified versions
could eventually be coupled with more conventional traffic simulation tools. In this
regard, the aim is to take the concept of driver model out of the research community
and bring it to a wider range of users, who can be engineers designing in-vehicle
systems and needing to understand the impact of their systems on driving or traffic
engineers needing to use driver models when changing a roadway design without
systematically having to conduct lengthy data collection.
The next step to pursue the development of a comprehensive driving model will
be to continue to take advantage of conferences to bring together model developers
and users in order to discuss about the elements to include on a driving model and
methods for exchanging the results of their research.
42 Cody and Gordon

Acknowledgments. We acknowledge the benefits we had from our exchanges with

the researchers who participated in the workshop. We also give thanks for the many
fruitful discussions we had with Jon Hankey, Dario Salvucci, Daniel Krajzewicz,
Mark Brackstone and Richard van der Horst during the preparation of the workshop
and with the audience on the day of the workshop.

References
Anderson, 1.R., Bothell, D., Byrne, M.D., Douglass, S., Lebiere, C. and Qin, Y. (2004) An
integrated theory of the mind. Psychological Review, 111, 1036-1060.
Brackstone, M. (2000). An examination of the use of fuzzy sets to describe relative speed.
Perception Ergonomics, 43(4), 528-42.
Brackstone, M., McDonald, M. and Wu, 1. (1997). Development of a fuzzy logic based
microscopic motorway simulation model. In Proceedings of the IEEE Conference on
Intelligent Transportation Systems (lTSC97). Boston, MA.
Jackson, D.L. and Bhise, V.D. (2002). An evaluation of the IVIS-DEMAnD driver attention
demand model. Report No. SAE 2002-01-0092, UMTRI-95608 A08. Human Factors in
Seating and Automotive Telematics, Warrendale, SAE, pp. 61-70.
Salvucci, D.D., Boer, E.R., and Liu, A. (2001). Toward an integrated model of driver
behavior in a cognitive architecture. Transportation Research Record, 1779, 9-16.
Salvucci, D.D., Zuber, M., Beregovaia, E., and Markley, D. (2005). Distract-R: Rapid
prototyping and evaluation of in-vehicle interfaces. In Proceedings ofHuman Factors in
Computing Systems. Portland, OR.
Theeuwes, 1. (2001). The effects of road design on driving. In Pierre-Emmanuel Barjonet
(Ed.). Traffic Psychology Today (pp. 241-263). Kluwer, Boston, MA.
Wu, 1., Brackstone, M. and McDonald, M. (2000). Fuzzy sets and systems for a motorway
microscopic simulation model. Fuzzy Sets and Systems 116(1), 65-76.
Wu, 1., Brackstone, M. and McDonald, M. (2003). The validation of a microscopic simula-
tion model: A methodological case study. Transportation Research C, 11(6),463-479.
3
Towards Monitoring and Modelling
for Situation-Adaptive Driver
Assist Systems
TOSHIYUKIINAGAKI

3.1 Introduction
In the classic tri-level study of the causes of traffic accidents, Treat et al. (1979) as-
cribe 92.6% of car accidents to human error, where human errors include improper
lookout (known as 'looking but not seeing'), inattention, internal distraction and
external distraction. Green (2003) reports that other studies have found similar
results that a human error is involved in 90% of car accidents. Human errors, such
as those listed in the above, can happen for everybody and may not be eradicated.
However, if there were some technology to detect driver's possibly risky behaviour
or state in a real-time manner, car accidents may be reduced effectively. Proactive
safety technology that detects driver's non-normative behaviour or state and pro-
vides the driver with appropriate support functions plays a key role in automotive
safety improvement. Various research projects have been conducted worldwide to
develop such technologies (see, e.g. Witt, 2003; Panou et al., 2005; Saad, 2005;
Amditis et al., 2005; Tango and Montanari, 2005; Cacciabue and Hollnagel, 2005).
This paper gives an overview on two of research projects in Japan, which aim
to develop technologies to detect driver's behaviour or state that is inappropriate
to a given traffic environment so that the driver may be provided with support for
enhancing his or her situation awareness or for reducing risk in the environment.
The first project is the 'Behaviour-Based Human Environment Creation Technol-
ogy', which was conducted during the period of 1999 to 2003, with the support of
the Ministry of Economy, Trade and Industry (METI), Government of Japan, and
the New Energy and Industrial Technology Development Organization (NEDO).
The other project is the' Situation and Intention Recognition for Risk Finding and
Avoidance' which has been proceeding since 2004, with the support of the Ministry
of Education, Culture, Sports, Science and Technology (MEXT), Government of
Japan.
This paper tries to focus on the modelling-related aspects of the projects, and
picks up a model from each project. From the first project, is the Bayesian network
model for detecting non-normative behaviour of the driver, in which the model
has been constructed based on driving behavioural data collected in the real traffic
environment. From the second project, is the discrete-event model of dynamical

43
44 Inagaki

changes of driver's psychological state, which has been developed to analyse and
determine how decision authority should be distributed between the driver and
automation under possibility of the driver's overtrust in 'smart and reliable' au-
tomation. The models are still in their early stages; however, they are expanding
description capabilities and applicabilities in the real world.

3.2 Behaviour-Based Human Environment Creation

Technology Project
3.2.1 Aims of the Project
When a driver's performance is suitable to task demands determined by traffic
situations (such as vehicle performance, vehicle speed, road structure, weather,
other traffic), the driver can enjoy safe driving. However, if the driver's performance
fails to adapt to the task demands, his or her behaviour may increase risk to the
situation (see Fig. 3.1). In order to evaluate the level of risk, traffic situations as
well as driving performance need to be sensed and monitored.
The Behaviour-Based Human Environment Creation Technology project was
conducted during the period of 1999 to 2003, with the support ofMETI and NEDO .
The project had four aspects, and one of the aspects was to create behaviour-
based driving assist technologies that determine whether a driver's performance
is deviated from the normative one in a given traffic situation as well as provide
the driver with a warning or advice, when appropriate. A driver's normative (or
baseline) behaviour was defined as his or her usual behaviour. The project took as an
object of study a driver who usually bears safe driving in mind and used SAS 592, a
self-rating scale of driving attitude , developed by the National Research Institute of
Police Science, Japan, to exclude inherently unsafe drivers from the data collection
phase .

Crash
(Fail to perfonn task)

FiGURE 3.1. Driving performance and task demand (Akamatsu and Sakaguchi , 2003); orig-
inally from Fuller and Santos (2002) .
3. Towards Monitoring and Modelling for Situation-Adaptive Driver Assist Systems 45

3.2.2 Measurement ofDriving Behaviour

In order to detect a driver's deviation from his or her normative performance,
good knowledge on the driver's usual performance is necessary. Vehicles with
a driving recorder system were developed, for the project, to record and collect
behavioural data in real road environments. The system consists of sensors, small
CCD cameras, a signal-processing device and a laptop on-board computer. The
recorded data include steering wheel angle, tum signal, wiper activation, strokes
of the brake and gas pedals, position of the right foot of the driver, geographical
position of the vehicle, velocity and state of the vehicle, and relative distance and
speed to the lead and following vehicles.
Eight different driving routes were chosen for investigation as 30-min trips
with several right and left turns. Subjects were recruited and asked to drive a
specified route once a day on weekdays for 2 months. Forty sets of behavioural
data were collected for each subject per route (Akamatsu et al., 2003). At the end
of data-collection phase, each subject was given the Driving Style Questionnaire
developed for quantifying driver's attitudes, that may be related to or may affect
his or her driving behaviour (Akamatsu, 2003). The total number of participants
was 92 (59 male and 33 female). The age ranged from 21 to 71 years for males
and from 20 to 66 years for females.

3.2.3 Driving Behaviour Modelling

A Bayesian network approach has been applied to driving behaviour modelling.
Bayesian networks (often called belief networks) are directed graph models, in
which an arc from a node, say A, to another, say B, represents a causal relation-
'A causes B'. The conditional probability distribution is specified at each node in
the network to represent how a child node is affected by its parent nodes.
For each situation of interest, a Bayesian network model can be developed.
Consider a case, for instance, in which a driver approaches an intersection with the
STOP sign. Behavioural events needed for describing the case include (a) release
of the gas pedal, (b) moving the foot to the brake pedal, (c) onset of braking,
(d) reaching the maximum deceleration and (e) full stop. Figure 3.2 depicts the
resulting Bayesian network model, obtained based on the collected behavioural
data. In the Bayesian network modelling, the data analysis results used were those
claiming that the weather conditions and the driver's score for the methodical scale
in the Driving Style Questionnaire can be regarded as performance-shaping factors
(Akamatsu and Sakaguchi, 2003).

3.2.4 Detection ofNon-Normative Behaviour

Once a Bayesian network model has been obtained for a driver's normative be-
haviour, it then becomes possible to compute the probability distribution for each
node in the network. Consider a case, as an example, in which some node, say
Z, denotes a random variable related to a safety-critical action for avoiding an
accident. Suppose Z has two nodes, X and Y, as its parents. Based on probability
46 Inagaki

FIG URE 3.2. Bayesian network model for the behavioural events when approaching an
intersection with the STOP sign (Akamatsu and Sakaguchi, 2003).

distributions for X and Y and the knowledge of probabilistic causal relation among
the nodes, one can derive the conditional probability distribution, P(ZIX , Y) for
each combination of values of X and Y. The conditional probability distribution
P(ZIX , Y) describes how Z can take different values when the driver behaves in
a normative manner.
Suppose an on-board sensor has observed that the random variable Z took a
value z at some time point. Hypothesis testing is then performed to determine
whether 'z may be regarded as a sample from the distribution P(Z IX, Y)'. If the
hypothesis was rejected at some level of significance, it is then concluded that
the driver's behaviour may be deviated from his or her normative performance.
Akamatsu et al. (2003) defined the ' level of normality' and have developed a
method to give a warning to the driver when the calculated level of normality
becomes less than a specified threshold value .

3.2.5 Estimation ofDriver's State

In the Behaviour-Based Human Environment Creation Technology project, meth-
ods for estimating driver's mental tension and fatigue were also investigated, be-
cause driver's status need s to be estimated to provide a driver with support functions
suitable to his or her operational capability.

3.2.5.1 Estimation of Driver 's Mental Tension

Mental tension is one of driver's internal factors affecting driving ability. Investiga-
tors in the project have found that chromogranin A (CgA) in saliva is a biochemical
3. Towards Monitoring and Modelling for Situation-Adaptive Driver Assist Systems 47

index of driver's mental strain (Sakakibara and Taguchi, 2003). The index was not
convenient for real-time sensing of driver's mental tension. They found, however,
that the steering operation and the head motion of a driver can replace salivary CgA
in estimating a level of driver's mental tension. Taguchi and his colleagues argued
that the steering operations at a frequency lower than 0.5 Hz may reflect a driver's
reduced activation state in a monotonous driving environment and that the driver's
intentional behaviour, such as a lane change, does not usually have influence on
operational performance at such low frequencies (Sakakibara and Taguchi, 2005;
Taguchi and Sakakibara, 2005). They also considered that while concentrating
on driving, a driver generally puts power into various muscles to control body
in response to vibrations from road surface, acceleration and/or deceleration of
vehicle, which may create stiffness in the shoulders. Experiments were conducted
to investigate relations between the two indices for tension. It has been observed
that the steering operation and the head motions are complementary to each other.
They have also found that a real-time estimation method can be implemented to
distinguish the four grades of driver's tension (viz. reduced activation, neutral,
moderate tension and hypertension).

3.2.5.2 Estimation of Driver's Fatigue

Fatigue is also one of the contributing factors affecting driving ability. Investigators
in the project have tried to develop a non-invasive sensing method for real-time
estimation of a driver's fatigue level (Furugori et al., 2003, 2005; Miura et al.,
2002). With sensor sheets, Furugori and his colleagues measured the pressure
distribution on the seat and found a relationship between changes in the load
centre position (LCP) calculated from the pressure distribution and driver-indicated
subjective fatigue. It was also found that the LCP contains information about
'prolonged changes' in posture with time (e.g. bending forward or backward) as
well as 'momentary changes' in posture (e.g. shifting weight). A 'postural change
parameter' and a 'weight-shifting parameter' were defined for the seatback and
the seat cushion. However, individual differences were observed in direction of
postural changes and weight shifting. A robust algorithm was thus required to
estimate fatigue level for each individual. Based on the ratios that indicate the
degree to which each parameter range deviates from those in normal driving, the
'fatigue index' was calculated to estimate a driver's fatigue level. It was found that
in around 90% cases the calculated fatigue index agreed with subjective fatigue
(Furugori et al., 2005).

3.3 Situation and Intention Recognition for Risk Finding

and Avoidance Project
3.3.1 Aims of the Project
Driving requires a continuous process of perception, decision and action. Under-
standing of the current situation determines what action needs to be done (Hollnagel
48 Inagaki

traffic environment

I perception I0::::::::::::::::;> L...---r------J c==;=~:.L.~~J

infer "intent"

"Is the driver's interpretation appropriate?"

Detect: mismatch ofbehavior and environment

FiGURE 3.3. Assessing appropriateness of a driver's situation recognition.

and Bye, 2000). In reality, however, drivers' situation recognition may not always
be perfect. Decisions and actions that follow poor or imperfect situation recogni-
tion can never be appropriate to given situations. It is not possible to 'see' directly
whether a driver's situation recognition is correct or not. However, monitoring the
driver 's action (or its precursor) and traffic environment may make it be possible
to estimate (a) whether the driver may have lost situation awarenes s, (b) whether
the driver's interpretation of the traffic environment is appropriate and (c) whether
the driver is inactive psychologically, e.g. due to complacency, or physiologically,
e.g . due to fatigue, (see Fig. 3.3).
Since 2005, the author has been conducting a 3-year research project , enti-
tled ' Situation and Intention Recognition for Risk Finding and Avoidance', with
the support of the MEXT. The aim of the project is to develop proactive safety
technologies to (a) monitor driver behaviour for assessing his or her intention, (b)
detect mismatches of traffic environment and driver's interpretation of it, (c) assess
the driver state and (d) provide the driver with appropriate assist functions in a
situation-adaptive and context-dependent manner (see Fig. 3.4).
The research topics in the project are categorised as follows : (1) estimation of
driver 's psychological and physiological state, (2) driver behaviour modelling, (3)
intelligent information processing for situation recognition and visual enhance-
ment and (4) adaptive function allocation between drivers and automation. In (I),
real-time methods are under development for detecting the driver's inattentive-
ness , hypo-vigilance, and complacency, through monitoring parameters, such as
body movement, dynamical changes of the LCP on the back, eye and head move-
ments , blinks , instep position of the right leg, operation of the steering wheel
and movements of gas and brake pedals (Itoh et al., 2006) . In (1), levels of
 w
~~ .:_ =-~ ~ ~. ,
'.--:--. . . -::4-- _.., ~ -. Q3
:E
~
- ''' , . .,
- -. '"a.
'"
~
o
::l
s"
:::!.
Driver's behaviors ::l
oe
::l
'"0-
I ~
o
0-
~
s
{Jt>

..,0'
co
;:
;:;
o
::l
:i>
0-
-g
'"
<.
rt>

.o.,
Risk finding <.
rt>
..,
:>
'"
'"
~"
(/)
'<
o
n;
Human -centered accident-prevention technology for transportation safety 3
'"
.j:>.
FIGURE 3.4. Situation and intention recognition for risk finding and avoidance project. \Q
50 Inagaki

driver's fatigue and drowsiness are also estimated by applying a chaos theoretic
method (Shiomi and Hirose, 2000) to a driver's voice during verbal communi-
cation. Driver modelling in (2) adopts a Bayesian network approach, as in the
case of the Behaviour-Based Human Environment Creation Technology project.
Some mathematical and information processing methods are under development
in (3) for machine learning, recognition of traffic environments and human vision
enhancement.
The methods in (1) to (3) give messages (or warnings) to the driver, when
they determine that the driver's situation recognition and intention may not be
suitable to a given traffic condition. If the driver responds quickly to the messages,
the potential risk shall be diminished successfully. If the driver fails to accept
or respond to the messages in a timely manner, on the other hand, accidental or
incidental risks may grow. Research topics in (4) investigate such situations. They
aim to develop an adaptive automation that can support drivers at various levels of
automation, which shall be discussed in the next section.

3.3.2 Adaptive Function Allocation Between Drivers

and Automation
A scheme that modifies function allocation between human and machine dynam-
ically depending on situations is called an adaptive function allocation. The au-
tomation that operates under an adaptive function allocation is called adaptive
automation (see e.g. Inagaki, 2003; Scerbo, 1996). Adaptive automation assumes
criteria to determine whether functions have to be reallocated, how and when.
The criteria reflect various factors, such as changes in the operating environment,
loads or demands to the operators and performance of the operators. Adaptive au-
tomation is expected to improve comfort and safety of human-machine systems in
transportation. However, it is known that the humans working with highly intelli-
gent and autonomous machines often suffer negative consequences of automation,
such as the out-of-the-loop familiarity problem, loss of situation awareness, au-
tomation surprises. When carelessly designed, adaptive automation may face with
such undesirable consequences. One of critical design issues in adaptive automa-
tion is decision authority over automation invocation (viz. who makes decisions
concerning when and how function allocation must be altered). The decision au-
thority issue can be discussed in a domain-dependent manner. Automobile is one
of domains in which machine-initiated control over automation invocation may be
allowed for assuring safety (Inagaki, 2006). Following are two examples that are
under investigation in the project.
Example 1: Suppose that the driver of the host vehicle H determines to make a
lane change because the lead vehicle A has been driving rather slowly. When the
host vehicle's on-board computer noticed that the driver glanced the side mirror
several times, it inferred that the driver had formed an intention of changing lane,
where the computer regarded 'glancing the side mirror several times in a short
period of time' as precursor to the action of changing lane (see Fig. 3.3). The
computer, monitoring backward with a camera, also noticed that a faster vehicle
3.Towards Monitoring and Modelling for Situation-Adaptive Driver AssistSystems 51

A H

B c D

FIGURE 3.5. An examplein which machine intelligence is given decision authority.

C is coming from behind on the left lane . Based on the understanding of driver's
intention and the approach of vehicle C, the computer puts its safety control
function into its armed position in preparation for a case when the driver chooses
a wrong timing to execute an action (viz. steering the wheel) due to improper
interpretation of the traffic environment. Now the driver, who has seen that a
very fast vehicle B almost passed him on the left, begins to steer the wheel to the
left, failing to notice vehicle C (Fig. 3.5). The computer immediately activates
the safety control function to make the wheel either slightly heavy to steer (soft
protection) or extremely heavy to steer (hard protection). The soft protection is
for correcting the driver 's interpretation of the traffic environment, and the hard
protection is for preventing a collision from occurring. The computer takes the
steering authority from the driver partially in cases of soft protection and fully
in cases of hard protection.
Example 2: Suppose that the driver of the host vehicle H wants to make a lane
change to the left, because the lead vehicle A drives rather slowly. When glancing
at the rear view mirror, the driver noticed that faster vehicles, C and D, are
coming from behind on the left lane (Fig . 3.5). By taking several looks at the
side mirror, the driver tries to find a precise timing to cut in. In the meantime,
based on the observation that the driver looked away many times in a short
period of time , the on-board computer determined that the driver has formed an
intention of changing lane and that he might not be able to pay full attention
to the lead vehicle A. The computer then puts its safety control function into
its armed position in preparation for a deceleration of the lead vehicle A. If the
lead vehicle A did not make any deceleration before the host vehicle's driver
completes a lane change, the computer will never activate the safety control
function and will put it back into a normal standby position. On the other hand,
if the computer detected a rapid deceleration of the lead vehicle A while the
driver is still looking for a timing to make a lane change, it immediately activates
its safety control function, such as an automatic emergency brake .

3.3.3 Decision Authority and the Levels ofAutomation

For the discussion of decision authority, the notion of the level ofautomation (LOA)
is useful. Table 3.1 gives an expanded version in which an LOA comes between
levels 6 and 7 in the original list by Sheridan (1992). The added level, called
52 Inagaki

TABLE 3.1. Scales of levels of automation.

1. The computer offers no assistance; human must do it all.
2. The computer offers a complete set of action alternatives and
3. narrows the selection down to a few, or
4. suggests one and
5. executes that suggestion if the human approves, or
6. allows the human a restricted time to veto before automatic execution, or
6.5 executes automatically upon telling the human what it is going to do, or
7. executes automatically, then necessarily informs humans,
8. informs him after execution only if he asks,
9. informs him after execution if it, the computer, decides to.
10. The computer decides everything and acts autonomously, ignoring the human.

After Sheridan (1992), Inagaki et al. (1998) and Inagaki and Furukawa (2004).

level 6.5, was first introduced in Inagaki et al. (1998) with twofold objectives:
(1) to avoid automation surprises possibly induced by automatic actions and (2)
to implement actions indispensable to assure systems safety in emergency. When
the LOA is positioned at level 6 or higher, the human may not be in command.
Generally speaking, it would be desirable, philosophically and practically, that
human is maintained as the final authority over the automation. However, as can
be seen in Examples 1 and 2, there are cases in which automation may be given
decision authority (Inagaki, 2006).

3.3.4 Model-Based Evaluation ofLevels ofAutomation

This section gives a driver model for evaluating design of interactions between a
driver and automation under the possibility of the driver's overtrust in the automa-
tion. Let us take a case, as an example, in which an adaptive cruise control (ACC)
system is available on the host vehicle. The ACC system is a partial automation for
longitudinal control, designed to reduce the driver's workload by freeing the driver
from frequent acceleration and deceleration. It controls the host vehicle so that it
can follow a vehicle ahead (the target vehicle) at a specified distance. When the
ACC system detects the deceleration of the target vehicle, it slows down the host
vehicle at some deceleration rate. As long as the deceleration of the target vehicle
stays within a certain range, the ACC system can control the speed of the host
vehicle perfectly and no rear-end collision into the target vehicle occurs. It would
be natural for the driver to trust in the ACC system while observing it behaves
correctly and appropriately. Sometimes the driver may place excessive trust in the
automation and may fail to allocate his or her attention to the traffic environment.
If the target vehicle makes a rapid deceleration at a high rate, the ordinary brake
by the ACC system may not be powerful enough to avoid a collision into the target
vehicle. In such cases, the automation issues an 'emergency-braking alert', which
tells the driver to hit the brake himself or herself hard enough to avoid a collision.
If the driver has been in a hypovigilant state, however, he or she may fail to respond
quickly to the situation.
3.Towards Monitoring and Modelling for Situation-Adaptive Driver AssistSystems 53

t
State IV: Hyper-normal and excited state. Target vehicle ~
approaches

C~:::>
State III: Normal and vigilant state
(the best state for safe driving).
emerge.ncy
-brakmg
State II+: Normal and relaxed state with
moderate level of trust in the
t , ! alert
automation.
~ CJL:)
State II : Normal and relaxed state, with experiencing
complete faith in the automation.

State I: Subnormal and inactive state.

perfect control
consecutively CD
FIGURE 3.6. Drivers' psychological state and their dynamic transitions.

3.3.4.1 Drivers' Psychological States and Their Transitions

Inagaki and Furukawa (2004) have distinguished five psychological states for
drivers, by modifying the original model by Hashimoto (1984):
State I: Subnormal and inactive state.
State II: Normal and relaxed state, with complete faith in the automation.
State 11+: Normal and relaxed state with moderate level of trust in the automation.
State III: Normal and vigilant state, which is the best state for safe driving.
State IV: Hyper-normal and excited state.
State transitions occur dynamically in time as time passes by. Suppose a driver's
psychological state was positioned initially at State III, when he or she started
driving . If the driver has observed a certain number of the ACC system's perfect
longitudinal controls in response to decelerations of the lead vehicle , the driver's
psychological state goes from III to II+. Observing some more consecutive perfect
controls by the ACC system , the driver's psychological state may change further
from State 11+ to II. If the driver felt alarm, while observing the host vehicle came
close to the lead vehicle, his or her trust in the ACC system goes down a bit, and
a state transition occurs, say, from State II to II+ or from State II+ to III (see
Fig. 3.6) .

3.3.4 .2 Driver's Response to an Alert

Each psychological state is characterised by a corresponding driver performance
in situation recognition and swiftness of response to an emergency-braking alert .
A model for the driver 's response to an emergency-braking alert may be given as
follows : (I) If the driver was in State I when the alert was set off, he or she does
not respond to the alert at all. (2) If the driver was in State II, he or she stays in
the same state with probability p and hits the brake pedal in T2 seconds . With
probability 1 - p, the driver state jumps into State IV. (3) If the driver was either
54 Inagaki

in State II+ or III, he or she applies the emergency brake himself or herself either
in T2+ or T3 seconds, respectively. (4) In State IV, the driver panics and fails to
take any meaningful actions to attain car safety. T2, T2+ and T3 are treated as
random variables with different means.

3.3.4.3 Evaluation of Efficacy of Levels of Automation

The following are realistically feasible design alternatives for cases in which the
target vehicle makes a rapid deceleration at a rate much greater than the maximum
deceleration rate which the ACC system can handle with its ordinary automatic
brake:
Scheme 1: Upon recognition of a rapid deceleration of the target vehicle, the ACC
system gives an emergency-braking alert, where the LOA is positioned at 4.
Scheme 2: Upon recognition of a rapid deceleration of the target vehicle, the ACC
system gives an emergency-braking alert. If the driver does not respond within
a pre-specified time (due to inattentiveness or delay in situational recognition),
it applies an automatic emergency brake, where the LOA is positioned at 6.
Scheme 3: Upon recognition of a rapid deceleration of the target vehicle, the ACC
system applies its automatic emergency brake simultaneously when it issues an
emergency-braking alert, where the LOA of this scheme is positioned at 6.5.
With the models of driver's psychological state and response time for each state,
Monte Carlo simulations can be performed to evaluate efficacy of the LOAs under
possibility of driver's overtrust in the ACC system. The following are some of
observations obtained (Inagaki and Furukawa, 2004):
(1) A safety control scheme with LOA-6 may not be effective, compared to one
with LOA-4, if the driver trusts in the automation excessively. In order to
mitigate the drawback of the LOA-6, some measures are needed to keep the
driver alert.
(2) The drawback of LOA-6 is partially due to the time delay of the automatic
safety control action. The LOA-6.5 scheme may be more effective than LOA-6
or LOA-4 in order to assure systems safety in time - criticality under possibility
of driver's complacency.

3.4 Concluding Remarks

This paper has given brief descriptions on two driver behaviour modelling re-
lated research projects in Japan the aims of which are to develop proactive safety
technologies to detect mismatches between driver's behaviours and the traffic en-
vironment. This paper discussed two models in the projects: one is the Bayesian
network model for detecting non-normative behaviour of the driver and the other
is the discrete-event model for analysing driver interactions with 'smart and reli-
able' automation. These models may still be in their early stages of development,
3. Towards Monitoring and Modelling for Situation-Adaptive Driver Assist Systems 55

compared to control theoretic models such as those discussed by Weir and Chao
(2005) and Juergensohn (2005). However, the models are expanding their descrip-
tion capabilities and applicability.
Drivers must be provided with necessary and sufficient supports by machine
or automation. Intention understanding and communication play important roles
in realising such meaningful support functions. If the driver fails to understand
the intention of machine intelligence, an automation surprise may happen. If the
machine does not recognise the driver's intention, its 'good support' may be an-
noying to the driver. The topic of intent inference attracts keen interests in Japan as
well as the rest of the world. Although no discussion could be made in this paper,
a fuzzy association system model with case-based reasoning has been developed
for recognising human intention through monitoring his or her behaviour in the
'Humatronics' project in Japan for the safety of drivers and pedestrians (Umeda
et al., 2005; Yamaguchi et al., 2004).
In spite of its usefulness, an intention understanding approach may have some
limitations, at least at the present time. The second project described in this pa-
per, for instance, sometimes found difficulty in understanding the intention of the
driver from his or her behaviour. Norman (1988) has distinguished seven stages
of action, in which the first-four stages are (1) forming the goal, (2) forming the
intention, (3) specifying an action and (4) executing the action. The project tries to
infer the intention of the driver by catching some 'precursor' to the action (recall,
Fig. 3.3) that may come between stages (3) and (4). Some of problems observed
in the project are as follows: (a) No useful precursor may exist for some intended
actions, which might also be individual-dependent. (b) The driver may form a goal
for a very immediate future in dynamically changing environment, which makes
it hard either to infer its associated intention or to identify in a timely manner a
driver support that is appropriate for the intention. (c) When the driver's action
was corrected by the machine, the driver is likely to perceive that the machine did
not understand his or her intention at all, although the action needed to be cor-
rected because its execution timing did not match the traffic environment. Proactive
safety technologies usually assume machine intelligence. Failure in mutual under-
standing of intentions between the driver and the machine intelligence may bring
various inconveniences, such as automation surprises, distrust and overtrust. The
second project is now trying to challenge these problems via a context-dependent
adjustment of LOA as well as designing human interface that enables the driver
to (a) understand the rationale why the automation thinks so, (2) recognise in-
tention of the automation, (3) share the situation recognition with the automa-
tion and (4) perceive limitations of automation's functional abilities (Inagaki,
2006).

Acknowledgments. This study was conducted as part of the government project,

'Situation and Intention Recognition for Risk Finding and Avoidance' with the
support of the Ministry of Education, Culture, Sports, Science and Technology,
56 Inagaki

Government of Japan. The authorexpresses his appreciation to Dr.Cacciabue and

anonymous refereesfor theirconstructive comments and suggestions. Theyhelped
a lot in improving the quality and clarity of the paper.

References
Akamatsu, M. (2003). Measurement technologies for driver and driving behavior. In Pro-
ceedings oflEA 2003 (CD-ROM).
Akamatsu, M. and Sakaguchi, Y. (2003). Personal fitting driver assistance system based on
driving behavior model. In Proceedings of the lEA 2003 (CD-ROM).
Akamatsu, M., Sakaguchi, Y. and Okuwa, M. (2003). Modeling of driving behavior when
approaching an intersection based on measured behavioral data on an actual road. In
Proceedings of the Human Factors and Ergonomics Society 47th Annual Meeting (CO-
RaM).
Amiditis, A., Lentziou, Z., Polychronopoulos, A., Bolovinou, A. and Bekiaris, E. (2005).
Real time traffic and environment monitoring for automotive applications. In L. Macchi,
C. Re and P.C.Cacciabue (Eds.). Proceedings ofthe International Workshop on Modelling
Driver Behaviour in Automotive Environments (pp. 125-131).
Cacciabue, P.C. and Hollnagel, E. (2005). Modelling driving performance: A review of
criteria, variables and parameters. In L. Macchi, C. Re and P.C. Cacciabue (Eds.). Pro-
ceedings of the International Workshop on Modelling Driver Behaviour in Automotive
Environments (pp. 185-196).
Fuller, R. and Santos, J. (2002). Psychology and the highway engineer. In R. Fuller
and J. Santos (Eds.). Human Factors for Highway Engineers (pp. 1-10). Pergamon,
Amsterdam.
Furugori, S., Yoshizawa, N., Iname, C. and Miura, Y. (2003). Measurement of driver's
fatigue based on driver's postural change. In Proceedings ofthe SICE Annual Conference
(pp. 1138-1143).
Furugori, S., Yoshizawa, N., Iname, C. and Miura, Y. (2005). Estimation of driver fatigue
by pressure distribution on seat in long term driving. Review ofAutomotive Engineering,
26(1), 53-58.
Green, M. (2003). What causes 90% of all automobile accidents? Available at
http://www.visualexpert.com/accidentcause.html.
Hashimoto, K. (1984). Safe Human Engineering. Chuo Rodo Saigai Boushi Kyokai (in
Japanese).
Hollnagel, E. and Bye, A. (2000). Principles for modeling function allocation. International
Journal ofHuman-Computer Studies, 52, 253-265.
Inagaki, T. (2003). Adaptive automation: Sharing and trading of control. In E. Hollnagel
(Ed.). Handbook of Cognitive Task Design (pp. 147-169). Lawrence Erlbaum, Mahwah,
NJ.
Inagaki, T. (2006). Design of human-machine interactions in light of domain-dependence
of human-centered automation. Cognition, Technology and Work, 8,161-167.
Inagaki, T. and Furukawa, H. (2004). Computer simulation for the design of authority in
the adaptive cruise control systems under possibility of driver's over-trust in automation.
In Proceedings of the IEEE SMC Conference (pp. 3932-3937).
Inagaki, T., Moray, N. and Itoh, M. (1998). Trust self-confidence and authority in human-
machine systems. In Proceedings of the IFAC Man-Machine Systems (pp. 431-436).
3. Towards Monitoring and Modelling for Situation-Adaptive Driver Assist Systems 57

Itoh, M., Akiyama, T. and Inagaki, T. (2006). Driver behavior modeling. Part II: Detec-
tion of driver's inattentiveness under distracting conditions. In Proceedings of the DSC-
Asia/Pacific (CD-ROM).
Juergensohn, T. (2005). Control theory models of the driver. In L. Macchi, C. Re and
P.C. Cacciabue (Eds.). Proceedings of the International Workshop on Modelling Driver
Behaviour in Automotive Environments (pp. 37-42).
Miura, Y., Yoshizawa, N. and Furugori, S. (2002). Individual variation analysis of body
pressure distribution for long term driving simulation task. The Japanese Journal of
Ergonomics, 38(Suppl), 70-72.
Norman, D.A. (1988). The Psychology of Everyday Things. Basic Books, New York.
Panou, M., Bekiaris, E. and Papakostopoulos, V.(2005). Modeling driver behavior in EU and
international projects. In L. Macchi, C. Re and P.C. Cacciabue (Eds.). In Proceedings of
the International Workshop on Modelling Driver Behaviour in Automotive Environments
(pp.5-21).
Saad, F. (2005). Studying behavioural adaptations to new driver support systems. In L.
Macchi, C. Re and P.C. Cacciabue (Eds.). Proceedings of the International Workshop on
Modelling Driver Behaviour in Automotive Environments (pp. 63-73).
Sakakibara, K. and Taguchi, T. (2003). Biochemical measurement for driver's state based
on salivary components. In Proceedings of the lEA 2003 (CD-ROM).
Sakakibara, K. and Taguchi, T. (2005). Evaluation of driver's state of tension. In Proceedings
of the HCI International (CD-ROM).
Scerbo, M.W. (1996). Theoretical perspectives on adaptive automation. In R. Parasuraman
and M. Mouloua (Eds.). Automation and Human Performance (pp. 37-63). Lawrence
Erlbaum, Mahwah, NJ.
Sheridan, T.B. (1992). Telerobotics, Automation, and Human Supervisory Control. MIT
Press, Cambridge, MA.
Shiomi, K. and Hirose, S. (2000). Fatigue and drowsiness predictor for pilots and air traffic
controllers. In Proceedings of the 45th Annual ACTA Conference (pp. 1-4).
Taguchi, T. and Sakakibara, K. (2005). Evaluation of driver's state of tension. Review of
Automotive Engineering, 26(2), 201-206.
Tango, F. and Montanari, R. (2005). Modeling traffic and real situations. In L. Macchi, C.
Re and P.C. Cacciabue (Eds.). Proceedings of the International Workshop on Modelling
Driver Behaviour in Automotive Environments (pp. 133-147).
Treat, J.R., Tumbas, N.S., McDonald, S.T., Shinar, D., Hume, R.D., Mayer, R.E., Stansifer,
R.L. and Castellan, N.J. (1979). Tri-level study of the causes of traffic accidents: Final
report volume 1. Technical report, Federal Highway Administration, US DOT.
Umeda, M., Yamaguchi, T. and Ohashi, K. (2005). Intention recognition system using
case-based reasoning. In Proceedings of the RISP International Workshop on Nonlinear
Circuit and Signal Processing (pp. 243-246).
Weir, D.H. and Chao, K.C. (2005). Review of control theory models for directional and speed
control. In L. Macchi, C. Re and P.C. Cacciabue (Eds.). Proceedings ofthe International
Workshop on Modelling Driver Behaviour in Automotive Environments (pp. 25-36).
Witt, G.J. (2003). SAfety VEhicle(s) Using Adaptive Interface Technology (SAVE-IT) Pro-
gram, DTRS57-02-R-20003. U.S. Department of Transportation.
Yamaguchi, T., Matsuda, S., Ohashi, K., Ayama, M. and Harashima, F. (2004). Humane
automotive system using driver and pedestrian intention recognition. In Proceedings of
the World Congress on Intelligent Transport Systems (CD-ROM).
 II
Conceptual Framework and
Modelling Architectures
4
A General Conceptual Framework for
Modelling Behavioural Effects of
Driver Support Functions
lORAN ENGSTROM AND ERIK HOLLNAGEL

4.1 Introduction
In recent years, the number of in-vehicle functions interacting with the driver
has increased rapidly. This includes both driving support functions (e.g. anti-lock
brakes, collision warning systems, adaptive cruise control) and functions support-
ing non-driving tasks, e.g. communication and entertainment functions. Today,
many of these functions are also featured on portable computing systems, com-
monly referred to as nomadic devices. Moreover, in order to handle this growth in
diversity and complexity of in-vehicle functionality, several types of meta func-
tions for human-machine interface integration and adaptation have been proposed.
Such functions, often referred to as workload managementfunctions, are intended
to resolve potential conflicts between individual functions with respect to their
interaction with the driver (see Engstrom et al., 2004; Brostrom et al., 2006).
The term driver support functions will henceforth be used to refer to in-vehicle
functions that support what drivers do, whether related to driving or not.
The proliferation of driver support functions naturally changes the nature of
driving and the different types of functions may induce a variety of behavioural
effects. One general reaction to new driver support functions is that drivers change
their behaviour in various ways to incorporate the functions into the driving task,
an effect commonly referred to as behavioural adaptation (Smiley, 2000; Saad
et al., 2004). Another issue that has received much recent attention is the effects of
multitasking while driving, e.g. driver distraction due to interaction with in-vehicle
functions, passengers or other objects in the vehicle.
While a wide range of driver models exists, addressing different aspects of driver
behaviour, a generally agreed conceptual framework for describing behavioural ef-
fects of driver support functions is still lacking. Technological and methodological
development in this area is therefore generally made without a common concep-
tual basis. The objective of this paper is to outline some basic requirements for
such a conceptual framework as well as to propose a specific candidate. The main
starting point will be models based on the cognitive systems engineering tradition,
specifically the COCOMIECOM framework (Hollnagel et al., 2003; Hollnagel and
Woods, 2005). It should be stressed that the aim here is not to present a validated

61
62 Engstrom and Hollnagel

model of driver behaviour but rather to propose a general conceptual framework

that can be used to describe behavioural effects of driver support functions, and
how these effects relate to accident risk. A principal motivation for this work was
the need for a common conceptual framework in the AIDE (Adaptive Integrated
Driver-vehicle Interface) ED-funded project, which deals with technical and hu-
man factor issues related to driver behaviour and automotive human-machine
interface development (see Engstrom et al. (2004) for an overview of the AIDE
project).
The chapter is organised as follows: In the next section, the general intended ap-
plication areas and the key requirements for the framework are outlined. Section 4.3
provides a review of existing driver behaviour models that are relevant for present
purposes. In Section 4.4, the key elements of the proposed framework are de-
scribed. In Section 4.5, some specific example applications are described. Finally,
Section 4.6 provides a general discussion and conclusions.

4.2 Intended Application Areas and Requirements

In order to derive the basic requirements for the intended conceptual framework, it
is necessary to first consider how it is intended to be used. As mentioned above, the
main purpose of the framework is not a validated model of driver behaviour. Rather,
the idea is that the framework should provide a common language to describe key
issues related to behavioural effects of in-vehicle functions. More specifically, the
framework should be applicable to (at least) the following problems:

4.2.1 Functional Characterisation of Driver

Support Functions
As in many other fields, the development of driver support functions is still to a
large extent driven by technological possibilities rather than the actual needs of
the users. This often results in 'solutions that are looking for a problem' (Holl-
nagel, 2006). As a result, driver support functions are often described in terms of
the underlying technology, e.g. 'driver monitoring' and 'vehicle-to-vehicle com-
munication', while it is not always entirely clear what driver goals or task they
are intended to support. In order to link driver support functions to their potential
behavioural effects, it is useful to be able to characterise the different functions
with respect to their intended purpose, i.e. which goal(s) they support (including
driving- as well as non-driving-related goals). The present framework should be
able to provide such a functional taxonomy.

4.2.2 Coherent Description ofExpected Behavioural Effects

of Driver Support Functions
The framework should be applicable as a coherent conceptual basis to describe all
different types of behavioural effects of driver support functions, from the effect of
 4. A General Conceptual Framework for Modelling Behavioural Effects 63

multitasking while driving to short- and long-term behavioural adaptations such

as risk compensation and over-reliance on the support function.

4.2.3 Conceptualising Relations Between Behavioural

Effects and Road Safety
While a basic understanding of behavioural effects of driver support functions is
an important goal in itself, a further critical issue concerns the relation between
such effects and actual road safety, e.g. changes in incident and accident risk
(see Dingus (1995) for a good discussion on this topic). While not intended as
a general accident model, the framework proposed here should be applicable to
these problems as well.

4.2.4 Specific Requirements

From these three general intended application areas, some more specific require-
ments for the framework can be outlined. First, since different functions support
different aspects of driving, from low-level vehicle handling to high-level naviga-
tion and route planning, the framework must be able to account for behaviour on
all levels and, equally important, the relation between the different levels. Second,
related to the previous requirement, it is important that the same principles of
analysis are applied on all levels of description, i.e. the model should be recursive
(Cacciabue and Hollnagel, 2005). Finally, the framework must be able to account
for time, i.e. the dynamics of behaviour. This is particularly important to capture
the self-paced, adaptive, nature of driving.

4.3 Existing Models of Driver Behaviour

Driver behaviour modelling has a long history and a wide range of existing mod-
els address different aspects of driving. However, models specifically targeting
behavioural effects of driver support functions are rare. According to Michon
(1985), a general distinction could be made between taxonomic and functional
models. The former refers to descriptive models, or 'inventory of facts', without
an account of the interaction between the model components, including trait mod-
els (for example of driver accident proneness) and task analysis. The latter refers to
models that account for processes and/or interactions within the modelled system.
The focus of this review is on functional models.

4.3.1 Manual Control Models

The early driver modelling efforts focused mainly on control-theoretic models of
vehicle handling (e.g. Weir and McRuer, 1968). In these models, vehicle control
was generally modelled in terms of feedback control mechanisms with the goal
of minimising the difference between a reference (or target) state (e.g. the desired
heading angle) and the actual state. Later developments of these models included
64 Engstrom and Hollnagel

elements of feed forward control (McRuer et aI., 1977; Donges, 1978). While these
types of models are useful for modelling manual lateral and longitudinal vehicle
control in constrained situations, they do not capture higher level aspects of driving
such as decision making, planning or motivation.

4.3.2 Information Processing Models

During the past 40 years, the information processing paradigm, based on the digital
computer as the main metaphor, has dominated human factors and cognitive sci-
ence. The basic idea behind this paradigm is that human cognition can be modelled
as sequences of logically separated computational steps, including perception, de-
cision and response selection. Human attention and performance limitations are
then modelled in terms of limited capacity at these different stages (Moray, 1967;
Kahneman, 1973). The information processing paradigm has had a great influence
on theories of multiple task sharing, which are often used to understand the ef-
fects of interacting with a driver support system while driving. In particular, the
concept of mental workload, defined by de Ward (1996) as 'the specification of
the amount of information processing capacity that is used for task performance'
(p. 15), is directly based on the limited capacity metaphor. Most experimental work
in this area also involved dual task studies to investigate the level of interference
between different types of tasks, leading to the influential multiple resource theory
(Wickens, 2002). However, with a few exceptions (e.g. Shinar, 1992; Salvucci,
2001) information processing models have not been incorporated in more general
driver behaviour models. Whether this should be considered as a failure of the
automotive human factors trade, as proposed by Michon (1985), or as due to the
fact that 'early information processing models and their associated experimen-
tal techniques were incompatible with the requirements of complex tasks such
as driving' (Ranney, 1994), could be discussed. In any case, one shortcoming of
information processing models when applied in the driving domain is the basic
notion of the human as primarily a passive receiver of information, which makes
it difficult to account for drivers' active management of traffic situations, e.g. by
means of self-pacing and dynamic task allocation.

4.3.3 Motivational Models

Many existing models characterise driver behaviour in terms of dynamic regula-
tion of risk. These types of models are often referred to as motivational models
(Ranney, 1994). By contrast to information processing models, motivational
models emphasise the self-paced nature of driving and attempt to understand
the dynamical adaptation to varying driving conditions. The main differences
among existing motivational models concern the criteria proposed to govern the
adaptation. For example, the risk-homeostasis theory (Wilde, 1982) hypothesises
that drivers strive to maintain a constant level of accepted risk. By contrast, the
zero-risk theory (Summala, 1985; 1988) states that the driver aims to keep the
subjectively perceived risk at zero-level. A related account has been offered by
Fuller (1984), who suggests that drivers' behaviour is guided by threat avoidance.
 4. A General Conceptual Framework for Modelling Behavioural Effects 65

In more recent model, the same author proposes that the driver attempts to maintain
a certain level of task difficulty rather than a level of risk (Fuller, 2005; see also
Chapter 10 in this book). Finally, based on Damasio's concept of somatic markers
(Damasio, 1994), Vaa (Chapter 12 in this book) proposes that driver behaviour is
largely driven by emotional responses to risky situations.
Motivational models have been criticised for being too unspecific regarding
internal mechanisms and, as a result, being unable to generate testable hypotheses
(Michon, 1985). It could also be argued that the need to include motivational or
emotional aspects partly is an artefact of the limitations of information processing
models. Since 'cold' cognition automatically excluded 'hot' cognition (Abelson,
1963), something important was missing from these models. This is, however, not
an issue in models that pre-dates information processing models, such as Gibson
and Crooks (1938, see below).

4.3.4 Safety Margins

As described in the previous section, a key issue in understanding driver behaviour
concerns the criteria that drivers use as the reference for adaptation. While the
motivational models use qualitative criteria such as risk or task difficulty, other
models have attempted a more quantitative approach based on safety margins. A
key starting point here is that humans, in most situations, tend to act as satisfiers
(rather than optimisers), i.e. they generally do not put more effort into a task than
needed (Simon, 1955). Many driver models incorporate the concept of subjectively
chosen safety margins as key criteria for guiding driver behaviour. Probably the
first account of safety margins in driving was offered by Gibson and Crooks (1938),
who proposed that drivers aim to stay within a 'field of safe travel', which can be
conceptualised as 'tongues' stretching out in front of the vehicle, with their size
and form being determined by the time-to-contact to surrounding obstacles. More
recently, this concept has been developed into more concrete time-based safety
margins parameters. There is abundant evidence that time-to-object information is
used by humans and animals for guiding locomotion (Gibson, 1979). Lee (1976)
used the perceptual variable tau, representing time-to-contact in terms of optically
specified parameters, to model drivers' braking behaviour. In traffic research, time-
to-collision is often used as a driving performance metric (van der Horst, 1990; van
der Horst and Godthelp, 1989; Minderhoud and Bovy, 2001). The corresponding
metric for lateral control is time-to-line-crossing (TLC). Godthelp et al. (1984)
demonstrated that TLC correlates strongly with driver's self-chosen occlusion
time. Too small TLC values are thus strong indicators of violations of the driver's
subjectively chosen safety margins. For a general account of safety margins in
driving, see Nilsson (2001).
Summarising the models reviewed in the present and the previous section, there
seems to be a strong convergence towards the general idea that driver performance
could be understood in terms of adaptation governed by some type of safety mar-
gins, although these are conceptualised differently by different authors, e.g. in
terms of objective quantitative parameters such as TTC and TLC, or more con-
cepts such as perceived risk or task difficulty.
66 Engstrom and Hollnagel

4.3.5 Hierarchical Models

A common approach is to model the driving task as a set of hierarchically organised
sub-tasks. A general hierarchical account of human performance is Rasmussen's
three-level model, which proposes a distinction between knowledge-based, rule-
based and skill-based performance (Rasmussen, 1983). Knowledge-based perfor-
mance is mainly needed in situations that have not been encountered before and
thus requires conscious deliberation, while rule-based performance refers to the
application of learned rules. Finally, skill-based performance refers to automated
skills that do not require any cognitive processing.
There are also several examples of hierarchical models that describe the driv-
ing task, e.g. Alexander and Lunenfeld (1986) who characterise the driving tasks
in terms of three levels: navigation, guidance and control. Another example is
Michon's influential description of the driving task in terms of strategic, tactical and
operational levels (Michon, 1985). These models, which have been called 'second
generation motivational models' (Ranney, 1994), describe goals and motives on
different levels, from general trip planning (strategic), via obstacle avoidance (tac-
tical) to immediate vehicle control (operational). It is assumed that activities on the
different levels interact dynamically but none of these models say much about how
this is actually accomplished. Hale et al. (1990) proposed to combine Michon's and
Rasmussen's levels into a two-dimensional matrix where (Rasmussen's) driving
performance levels are mapped onto (Michon's) driving task levels. Such a repre-
sentation is useful to describe differences between experts and novices in terms of
the degree to which different tasks in the hierarchy are automated (Ranney, 1994).
In most of these hierarchical driving models the different levels are (more or less
explicitly) intended to represent levels of information processing. A different type
of hierarchical driver model is the Extended Control Model (ECOM) proposed by
Hollnagel et al. (2003), which puts a hierarchical description of driving-related
goals (similar to Michon's model) into a control theoretic framework. ECOM pro-
vides a representation of the performance (rather than information processing)
on different levels of the driving task and also offers an account of the dynamic
interactions between concurrent activities on the different levels. In the current
version, four layers are proposed: targeting, monitoring, regulating and tracking.
The model is an extension of the Contextual Control Model (COCOM; Hollnagel,
1993; Hollnagel and Woods, 2005), which provides a general account of the dy-
namical coupling between perception (or, more generally, situation assessment),
decision and action. The COCOM and ECOM models are further described in the
following section.

4.4 A Conceptual Framework

In this section, the proposed conceptual framework is outlined. The section starts
with a short discussion on the key concept of behaviour and a description of how
concepts from control theory can be used to model the dynamics of behaviour.
 4. A General Conceptual Framework for Modelling Behavioural Effects 67

The key elements of the proposed framework, the COCOM and ECOM models,
are then presented.

4.4.1 Driver Behaviour as Goal-Directed Activity

Interaction between the driver and vehicle functions can be viewed as an instance
of the more general notion of driver behaviour. Thus, it is important to define
more precisely what we mean by behaviour. As suggested, e.g. by Dennett (1987),
behaviour, as opposed to mere bodily movement, can be defined as goal-directed
activity. According to Dennett, we understand the activity of other people by
adopting what he calls the intentional stance. This entails ascribing intentions and
goals to other agents, which enables us to make predictions about their behaviour
(naturally, these ascriptions may not always correspond to actual intentions of
people).
Based on this notion, driver behaviour can be understood in terms of what is
required to accomplish a number of different goals while driving. In addition to
driving related goals (such as keeping within the lane or reaching a destination),
drivers may also be occupied with other goals that are only vaguely related, or
entirely unrelated, to vehicle operation (such as finding a track on an MP3 player).
Thus, driver behaviour could be defined as the general pursuit of driving- and
non-driving-related goals while driving. Consequently, the general role of driver
support functions is to support the driver in accomplishing these goals. Driving
can further be defined as the subset of driver behaviours aimed towards goals
associated with vehicle operation. The role of driving support functions (a subset
of driver support functions) is thus to support driving-related goals.
A distinction can also be made between goals that are permanent (such as avoid-
ing accidents, complying with the traffic code and avoiding risk) hence applicable
to more than one journey, persistent (i.e., valid for the duration of a single journey,
such as reaching the destination) and transient in the sense that they may come and
go during a journey (e.g., overtaking the car in front, reaching a cafeteria before
noon).

4.4.2 Dynamical Representation ofDriver Behaviour

In order to understand driver behaviour in terms of goal-directed activity, we need
a way to describe how drivers' goals are dynamically achieved and maintained.
Control is a useful concept for describing the dynamics of goal-directed behaviour,
in man as well as in machines. In general, control can be understood as the ability
to direct and manage the development of events (Hollnagel and Woods, 2005).
Controlling a process means that actions are determined by the aim to achieve a
consistent goal state (often called the reference or target value), e.g. by means of
countering effects of external disturbances. Control is thus closely related to order-
liness or predictability, i.e. a controlled system is orderly, stable and predictable
while a system that is out-of-control is disorderly, unstable and unpredictable.
68 Engstrom and Hollnagel

There are two basic forms of control: In feedback (or compensatory) control,
the controller performs corrective actions based on the deviation between a desired
outcome (the goal) and the actual state. The prototypical example of a feedback
control system is the thermostat. Another type of control is feed forward (or an-
ticipatory) control. In this case, control actions are based on predictions of future
states and, hence, proactive rather than reactive. Driving behaviour is generally a
mixture of feedback and feed forward control.
In engineering control applications, such as a thermostat or an Adaptive Cruise
Control system, the target values (i.e. the desired temperature and time gap re-
spectively) are generally determined beforehand by the engineer or set by the user.
However, when analysing human controlled behaviour, identifying the target, or
controlled variable, is often viewed as the key issue (Powers, 1998; Marken, 1986).
The indentification of the controlled variable is complicated by the fact that the
human controller, as mentioned above, seldom operates as an optimiser but rather
acts as a satisfier (Simon, 1955), thus tolerating a certain deviation from the target
value. In driving, time based safety margins (e.g. van der Horst and Godthelp,
1989) or risk thresholds (Summala, 1988) could be thought of as reference values
for the vehicle control loop (although this idea has not yet, to our knowledge, been
exploited in extisting models).
As described in the previous section, control theory has been widely applied
to the modelling of vehicle handling, in automotive and other domains (e.g. Weir
and McRuer, 1968; Donges, 1978). Control theory has also been applied to the
modelling of higher-level aspects of driving. For instance, many of the moti-
vational models reviewed above, especially the risk homeostasis model (Wilde,
1982), are based on control theoretical concepts. However, few existing models
allow for a unified representation of controlled behaviour at different levels of
the driving task. One existing modelling framework able to provide such descrip-
tions is the COCOMIECOM model (Hollnagel and Woods, 2005), which thus
has been selected as the main starting point for the proposed conceptual frame-
work. The next two sections describe the COCOM and ECOM models in more
detail.

4.4.3 The Contextual Control Model (COCOM)

The contextual control model (COCOM), described in Hollnagel (1993) and Holl-
nagel and Woods (2005), provides a general account for modelling human control
of a process or plant, based on Neisscr's (1976) perceptual cycle concept. An im-
portant starting point for COCOM is that the controller and controlled system is
viewed as a joint cognitive system (JCS). The central object of study is then the
JCS rather than the controller or the controlled system in isolation. This approach
offers a perspective that differs from traditional information processing models
where the human operator and the machine are normally treated as logically sep-
arate entities. The boundaries between the JCS and its environment are generally
determined by the objective of the analysis. Pragmatically speaking, an object is
 4. A General Conceptual Framework for Modelling Behavioural Effects 69

considered part of the JCS if (I) it is considered important for the ability of the
JCS to maintain control and (2) it can be controlled by the JCS. Moreover, objects
satisfying (2) but not (I) may be included in the JCS if this facilitates the purpose
of the analysis (Hollnagel and Woods, 2005). For present purposes, the JCS of
interest is in most cases the joint driver-vehicle system (JDVS).
COCOM is a general model of how control is maintained by a JCS and is appli-
cable across a range of different JCS types on different levels of description. This
type of analysis makes minimal assumptions about internal cognitive processes
and focuses on behaviour and the dynamical interactions between the compo-
nents in the JCS, in particular on how the JCS maintains, or loses, control of a
situation .
A central concept in COCOM is the construct-action-event cycle. The con-
struct refers to what the controller knows or assumes about the situation in which
the action takes place. The construct is the basis for selecting actions and in-
terpreting information. The selected actions affect the process/application to be
controlled. This generates events that provide feedback on the effects of the
action which, together with external disturbances, modifies the construct and,
hence, the future action selection . An important property of this model is thus
that it accounts for both the feedback and feedforward aspects of control, i.e. ac-
tion selection is a function of both direct feedback and the predictions of future
events.
The main factors determining the level of control maintained by a JCS is pre-
dictability and available time. A key property of COCOM is that it offers an
explicit account of time. Figure 4.1 gives an illustration of the three main temporal
parameters involved: (I) Time to evaluate events (T E) , (2) time to select an action
or response (T s) and (3) time to perform an action (T p ; see Hollnagel and Woods,

External event I
... ~ disturbance

Events l ...........
feedback
TE =time to
evaluate
events
Tp =time to
perform an
action

FiGURE 4.1. The contextual control model (Hollnagel and Woods, 2005) .
70 Engstrom and Hollnagel

2005, for a more elaborate description of the timing relations in COCOM). As

in any control system, the relations between these time parameters determine the
performance of the JCS. The relation between the time needed to evaluate, select
and perform an action (TE , Ts and Tp ) and the time available (TA ) is of special
importance. In order to maintain control, the time needed to perform one cycle
(TE + Ts + Tp ) must, in the long run, be less than the total time available. If the
available time becomes too small, it may be increased by slowing down the pace
of the task, e.g. by reducing speed in the case of driving (as long as the task is
self-paced). It is in many cases uncertain how much time is available, due to the
dynamics of the environment, including the unpredictability of other drivers. In
such cases humans tend to sacrifice thoroughness in order to maintain efficiency,
as described by the efficiency-thoroughness trade-off (ETTO) strategy (Hollnagel,
2004).

4.4.4 The Extended Control Model (ECOM)

While COCOM only describes a single control process (i.e. pursuit of a single
goal), the driving task generally involves the pursuit of several simultaneous sub-
goals with different time frames. A long-term goal could be to reach a destination
in time. An example of a medium-term scale goal is to overtake a vehicle ahead,
while short-term goals include staying in lane and avoiding obstacles. Note that
the temporal characteristics of goals is different from whether they are permanent,
persistent, or transient. A goal may also subsume goals on shorter time frames. For
example, in order to reach a destination in time, it may be necessary to overtake
a number of vehicles. This, in turn, requires safe vehicle handling in order to
avoid collisions. Thus, the driving task can be described as a set of simultaneous,
interrelated and layered control processes. In addition, drivers generally also pursue
goals that are unrelated to driving, e.g. talking to a passenger, using the cell phone,
looking for a place to eat, etc.
As reviewed above, this hierarchical organisation is reflected in many models of
driving, e.g. Alexander and Lunenfeld (1986) and Michon (1985). However, none
of these models provide a sufficient account of the dynamical aspects of driving
(i.e. the relation between performance and time), or the simultaneous relations
between control processes on different layers. This is offered by the Extended
Control Model (ECOM) (Hollnagel et al., 2003; Hollnagel and Woods, 2005),
which represents a multi-layered extension to COCOM. The basic structure of
ECOM is illustrated in Fig. 4.2.
A key assumption behind the ECOM model is that goals on different layers are
pursued simultaneously and that these goals and their associated control processes
interact in a non-trivial way. In the current version, four control layers are pro-
posed: Tracking, regulating, monitoring and targeting, where each layer potentially
contains multiple parallel control processes (it should be noted that these partic-
ular layers are not fixed and may be revised and adapted to different application
domains).
 4. A GeneralConceptual Framework for Modelling Behavioural Effects 71

Targeting ----

Anti cipatory
control

Tracking

FIGUR E 4.2. The extended control model (ECOM).

A key property of the model is that during normal (controlled) task performance,
the goals/targets for the control processes on a given layer are determined by the
control processes one layer up. In the driving domain, the tracking control refers to
the momentary, automated, corrections to disturbances, e.g. wind gusts . Regulat-
ing refers to more conscious processes of keeping desired safety margins to other
traffic elements. This determines the target values for the tracking control loops .
Monitoring refers to the control of the state of the joint vehicle-driver system rela-
tive to the driving environment. It involves monitoring the location and condition
of the vehicle, as well as different properties of the traffic environment, e.g. speed
limits. This generates the situation assessment that determines the reference for
the regulating layer. Finally, the targeting control level sets the general goals of
the driving task , which determines the objectives for the monitoring layer. The
functional characteristics of the four layers , in the context of driving, are described
in Table 4.1.
As illustrated in Fig. 4.2, tracking control is typically based on feedback (com-
pensatory control) while monitoring and targeting are mainly of the feed-forward
(anticipatory) type . The regulating layer may involve a mix of feedback and feed
forward control. The ECOM model provides an account of how goals at different
layers interact and how higher goals propagate all the way down to moment-to-
moment vehicle handling. Control tasks on different layers may also interfere with
each other and disturbances on lower layers may propagate upwards. For exam-
ple, looking for directions (monitoring) may disrupt visual feedback, which may
affect regulating and tracking control ; a sudden experience of slipperiness due to
ice on the road , manifested on the tracking level may modify higher-level goals
(e.g. increasing safety margins on regulating level) and even change the targets for
driving (e.g . choosing a different route) .
72 Engstrom and Hollnagel

TABLE 4.1. Functional characteristics of the EeOM layers.

Typeof control Demandsto Frequencyof Typicalduration
involved attention occurrence
Targeting Goal setting High, Low (mostly Short (minutes)
(feedforward) concentrated pre-journey)
Monitoring Condition Low (car). High Intermittentbut 10 min to duration
monitoring (traffic, regular (car). of voyage
(feedback + hazards) Continuous
feedforward) (traffic,
hazards)
Regulating Anticipatory High (uncommon Very high (town), 1 s - 1 min
(feedback + manoeuvres), medium
feedforward) Low (common (country)
manoeuvres)
Tracking Compensatory None Continuous <Is
(feedback) (pre-attentive)

4.5 Application
This section gives some examples of how the proposed framework can be applied
in the three problem domains stated in Section 2: (1) Characterisation of driver
support functions, (2) description of their behavioural effects and (3) reasoning
about the relation between behaviour/performance and road safety.

4.5.1 Characterising Driver Support Functions

In terms of the proposed framework, driver support functions are characterised
with respect to the goal(s) they are intended to support (where some functions
support non driving-related goals). Based on this, a tentative taxonomy of in-
vehicle functions could be outlined. The sub-set of driving support functions (with
'driving' defined as in Section 4.1) can be categorised with respect to the ECOM
layers described in the previous section. In addition, driver support functions
include non-driving related functions as well as workload management 'meta'
functions with the purpose to manage the different driving and non-driving related
functions with respect to their interaction with the driver. These function categories
are further described below, and illustrated in Fig. 4.3.

4.5.1.1 Support for Tracking

This category includes functions which support driving by (partly) automating
tracking control actions of speed and direction, e.g. antilock brakes (ABS), dy-
namic stability and traction control (DSTC), adaptive cruise control (ACC) and
lane keeping aid (LKA). For some functions, such as ACC, this means that human
involvement is to a large extent re-allocated to the regulating control layer. In fully
 4. A General Conceptual Framework for ModellingBehavioural Effects 73

Driver support functions

Driving support Non-driving

related functions

Targeting

Monitoring Speed alert

Regulating Co llision warning

Tracking ACC ABS

Worklo ad
management

FiG URE 4.3. Examplesofdriversupportfunctions mappedontotheECOMlayers. Workload

management can be viewedas a metafunctionthatarbitratesbetweendifferentcontroltasks.
(See the text for further explanation.)

automated driving, the driver's task would be limited to targeting, e.g. setting the
destination, perhaps supported by monitoring of non-instrumented information
sources.

4.5 .1.2 Support for Regulating

These are functions with the purpose to support regulatory control, for instance
helping the driver maintain adequate safety margins. This can be done , e.g. by
providing warnings when safety margins are about to be violated or otherwise
enhancing the perception of safety margins. Examples include forward collision
warning (FeW), lane departure warning (LDW) and night vision systems.

4.5.1 .3 Support for Monitoring

In-vehicle functions belonging to this category support the monitoring of higher
level aspects related to the driver-vehicle-environment state . Examples include
vehicle state monitoring functions (e.g. fuel and oil level indication), route guid-
ance and traffic information functions. Importantly, activities on the monitoring
level may interfere with tracking and regulatory control, as further discussed
below.

4.5 .1.4 Support for Targeting

Targeting is in driving, as well as in most other domains, predominantly a human

activity. To some extent a smart navigation system can be seen as impinging on
that , but on the whole, targeting goals are set by humans and not by machines.
74 Engstrom and Hollnagel

4.5.1.5 Non-Driving-Related Functions

In addition to the different types of driving support functions mentioned above,
there are many in-vehicle functions that primarily support other tasks than driving.
Examples include the radio, the media players, the phone and various functions
supporting the work task of professional drivers, e.g. fleet management functions.
However, it should be noted that such functions may be used to support driving,
e.g. in a situation where the mobile phone is used to obtain directions or a traffic
broadcast forces the driver to revise his/her goals. These functions are nevertheless
conceptually distinct from driving support functions by the fact that they are not
primarily intended to support driving.

4.5.1.6 Workload Management Functions

Workload management functions can be viewed as 'meta-functions', responsible
for coordinating individual functions, e.g. by means of information prioritisation
or by putting non-critical information on hold in demanding driving situations. So
far only a few systems of this type have entered the market (e.g. Volvo cars' intel-
ligent driver information system (IDIS) and Saab's dialogue manager), but more
advanced functions are being developed in different research efforts, both in-house
at the companies (see Brostrom et aI., 2006, for a description of the next generation
IDIS) and in collaborative efforts such as COMUNICAR (Amditis et aI., 2002),
AIDE (Engstrom et aI., 2004) and SAVE-IT (SAVE-IT, 2001). By contrast to the
individual functions described above, the workload management meta-functions
do not directly support specific driver goals. Rather, in terms of the proposed
framework, these functions could be conceptualised as indirectly supporting the
driver by resolving conflicts between different (driving- or non-driving related)
goals. One key objective is to promote safe driving by means of supporting the
driver in prioritising the tracking and regulating control tasks in demanding driving
situations. It should be noted that the adaptivity provided by workload manage-
ment functions, e.g. changing presentation format to resolve conflicts with other
functions or the driving situation, potentially reduces the predictability of system
behaviour. Since, as mentioned above, predictability is of key importance for the
JCS to maintain control, great care must be taken in the design of such functions
in order to avoid unexpected usability and/or safety problems.

4.5.2 Characterising Behavioural Effects of

Driver Support Functions
The behavioural effects of a specific driver support function is the result of a
complex dynamic interaction between individual driver characteristics (motivation
for driving, subjectively chosen safety margins, driving skills, personality, effort
etc.), vehicle parameters (e.g. steering and braking dynamics) and the driving
environment (road type, curvature, lane width, traffic density etc.). It is useful to
make a general distinction between direct and indirect (side) behavioural effects.
 4. A General Conceptual Framework for Modelling Behavioural Effects 75

Direct effects are those that are intended by the system designers and implied by the
system's functional specification. For driving support functions, direct effects are
normally the intended performance enhancements on one or more control layers
(e.g. increased lane keeping performance for LDW and increased route-finding
ability for navigation support). By contrast, indirect effects are not intended by
the designers (and thus not implied by the functional specification). It should be
noted that indirect effects are not necessarily bad. For example, a potential positive
indirect behavioural effect of a lane departure warning system could be an increased
use of the turn signal.
In the following sections, some of the main types of behavioural effects
found in the literature are discussed from the perspective of the proposed frame-
work.

4.5.2.1 Behavioural Adaptation to Driving Support Functions

The ability to adapt to changing circumstances is central to human (and animal)
behaviour. The term behavioural adaptation (BA) generally refers to 'the whole
set of behaviour changes that are designed to ensure a balance in relations between
the (human) organism and his surroundings and at the same time the mechanisms
and processes that underlie this phenomenon' (Bloch et al., 1999). However, in
traffic research, behavioural adaptation is often used to refer to a more specific
type of adaptation, as defined by OECD (1990): 'Those behaviours which may
occur following the introduction of changes to the road-vehicle-user system and
which were not intended by the initiators of the change.' (p. 23). Hence, accord-
ing to this definition, BA only refers to indirect effects, e.g. compensatory be-
haviours that may reduce, or even cancel out, the expected benefits of a safety
measure.
BA has been demonstrated for many different types of driver support functions
(see Smiley, 2000 and Saad et al, 2004, for examples). According to the proposed
framework, such changes can be more precisely described in terms of performance
on the ECOM control layers and the relation between them. For example, in one
of the most cited studies on BA, it was found that taxi drivers with ABS- (anti-lock
braking system) equipped-vehicles tended to drive with higher speed and adopt
shorter headways than drivers of vehicles without ABS (Fosser et aI., 1997). (This
particular effect was actually predicted almost 60 years earlier by Gibson and
Crooks (1938)). According to the proposed framework, this phenomenon can be
described in terms of a change of goals and criteria on the regulating layer, caused
by changing conditions on the tracking layer. When ABS is activated, drivers expe-
rience that the improved braking (i.e. longitudinal tracking) performance increases
their safety margins due to the shorter stopping distance. Thus, if there is a high-
level motivation (on the targeting level) for arriving quickly at the destination (as is
often the case for taxi drivers), the drivers can reduce headway and increase speed
while still keeping within their subjective safety margin. (Whether the actual risk
remains constant is, however, another issue). This safety margin threshold could
thus be viewed as the reference value for the tracking loop. This example shows
76 Engstrom and Hollnagel

how the dynamical interaction between high- and low level driving goals can be
conceptualised by the present framework.

4.5.2.2 Effects of Multitasking While Driving

An important issue that has attracted much recent attention is how the time sharing
between driving and other tasks affects driving performance and safety. A typical
case of such multitasking is the use of different in-vehicle information functions
while driving. These tasks are often referred to as secondary tasks while driving
is considered the primary task. The effects of secondary tasks are often concep-
tualised in terms of distraction, which has been defined as 'attention given to a
non-driving related activity, typically to the detriment of driving performance' (ISO
TC22/SCI3/WG8 CD 16673). However, this type of terminology creates concep-
tual difficulties because 'driving-related activity' and 'driving performance' are
not further defined. For instance, reading a map on the navigation system display
should clearly be regarded as a 'driving related activity'. On the other hand, it
could at the same time distract the driver and cause degraded driving performance.
Related problems have also been noted in analysis of naturalistic driving data in
the recent 100 car study. As pointed out by Neale et al. (2005):
'Historically, driver distraction has been typically discussed as a secondary
task engagement. Fatigue, has also been described as relating to driver inatten-
tion. In this study, it became clear that the definition of driver distraction needed
to be expanded to a more encompassing 'driver inattention' construct that in-
cludes secondary task engagement and fatigue as well as two new categories,
'Driving-related inattention to the forward roadway' and 'non-specific eye glance'.
'Driving-related inattention to the forward roadway' involves the driver checking
rear-view mirrors or their blind spots. This new category was added after viewing
multiple crashes, near crashes and incidents for which the driver was clearly pay-
ing attention to the driving task, but was not paying attention to the critical aspect
of the driving task (i.e. forward roadway) at an inopportune moment involving a
precipitating factor.' (p. 6).
The proposed framework reconciles these conceptual difficulties in a quite
straightforward way. As suggested above, the driving task should not be seen
as a single activity, but rather as a set of multiple simultaneous and layered control
tasks. Thus, driving performance could be defined with respect to any of these
control tasks. For example, driving performance on the tracking level is associated
with the ability to keep the vehicle within acceptable safety margins. Similarly,
performance on the regulating level would be related to the ability to select appro-
priate safety margins based on a general situation assessment at the monitoring
level. This situation assessment may induce inattention to the forward roadway
(e.g. when checking mirrors), and could hence be described as a distraction on the
tracking and/or regulating layers. In general, distraction with respect to a given
control process (e.g. tracking) could thus be viewed in terms of interference by
another (driving- or non-driving related) control process, typically resulting in
degraded performance on the given control task. Historically, 'driver distraction'
 4. A General Conceptual Framework for Modelling Behavioural Effects 77

often implicitly refers to interference with the tracking and/or regulating control
tasks (and this is probably the intended meaning of the definition cited above).
However, the present framework enables a more precise conceptualisation and
makes it possible to describe in more detail which tasks/control processes that are
affected in a given distraction scenario.
Another commonly used concept is mental workload. While distraction is de-
fined on the basis of attention allocation, driver mental workload refers to the
amount of resources that the driver needs to perform one or more tasks, relative
to a limited subjectively defined resource pool. As mentioned in section 4.3.2,
the concept has strong roots in the information processing paradigm (e.g. Moray,
1967). It should be noted that mental workload is not a necessary precondition
for distraction, since inappropriate attention allocation may be caused by low-
workload tasks as well, e.g. daydreaming, checking mirrors and looking at road
signs. One limitation of the traditional workload concept, based on the limited
capacity metaphor, is that it does not account well for the dynamics of self-paced
driving. Based on the COCOMIECOM model, a more dynamic view of workload,
viewing the driver as an active agent, can be outlined where the spare resources
for a control process can be viewed in terms of the difference between total time
available and the total time needed to perform the control loop (see Fig. 4.1). Thus,
for example, if the available time for the tracking control loop is reduced, e.g. due
to time sharing with another visually demanding task (e.g. entering a mobile phone
number) the driver can gain time by reducing speed. The mechanisms that drive
this type of adaptive behaviour can, again, be understood in terms of the balance
between higher-level goals (e.g. the desire to arrive in time, i.e. targeting) and
lower level goals (e.g. to keep acceptable safety margins, i.e. regulating).
In order to further illustrate the potential benefits of the proposed framework
compared to traditional information processing models, it is useful to take a closer
look at some empirical data from the HASTE EO-funded project. As part of
the project, a set of parallel experiments were conducted in different sites across
Europe, with the specific objective to investigate systematically the effects of visual
and cognitive load on driving performance and state (see, e.g. Engstrom et aI.,
2005a; Victor et aI., 2005; Jamson and Merat, 2005; Markkula and Engstrom, 2006;
Ostlund et al., 2004). In short, the HASTE results showed that visual time sharing
induced increased lane position variation, increased number of large steering wheel
reversals, reduced speed and increased headway to lead vehicles. By contrast, time
sharing with purely cognitive tasks (i.e. tasks that require no visual interaction)
did not interfere with tracking control at all, a result consistent with the meta-
analysis of mobile-phone studies made by Horrey and Wickens (2004). Rather,
results from HASTE indicate that cognitively loading tasks lead to significantly
improved tracking control in terms of reduced lane keeping variation compared to
baseline driving (e.g. Engstrom et aI., 2005a; Jamson and Merat, 2005), an effect
that has also been documented in other studies (e.g. Brookhuis et aI., 1991). This
increased lane keeping performance for cognitive tasks was also accompanied by
a concentration of gaze towards the road centre (Victor et aI., 2005), an effect also
found in previous studies (e.g. Harbluk and Noy, 2002; Recarte and Nunes, 2003)
78 Engstrom and Hollnagel

and an increased number of steering micro corrections (Markkula and Engstrom,

2006). Other studies have also found that cognitive load impairs signal detection
performance (e.g. Greenberg et al., 2003; Engstrom et al., 2005b). Finally, speed
adaptation is seldom observed for cognitive tasks (Ostlund et al., 2004) and some
of the HASTE experiments found that the longitudinal safety margin, in terms of
time-headway to a lead vehicle, was even reduced during cognitive load (Jamson
and Merat, 2005).
A limited capacity model, viewing workload as the single factor that modifies
behaviour (i.e. 'high workload leads to worse performance'), is of limited use
for understanding these behavioural patterns. For instance, it would be hard to
make sense of the fact that lane keeping variation is increased for visual tasks
but reduced for cognitive tasks. In terms of the present framework, the effects of
the visual task can be understood as a degradation of lateral tracking control due
to reduced visual input during gazes away from the road. However, this is com-
pensated for, at least to some extent, on the regulating layer, by means of speed
reduction. By contrast, cognitive load does not interfere with tracking, but rather
seems to reinforce the lateral tracking performance (for possible explanations for
this peculiar phenomenon, see Engstrom et al., 2005a and Victor, 2005). In terms
of ECOM, the main interference of cognitive tasks seems to be on the regulating
and monitoring layers, as evidenced by the reduced detection performance, the
inability to adapt speed and the reduced time headway. It should be stressed that
the COCOMIECOM framework, as applied here, does not offer a detailed expla-
nation of the mechanisms underlying these phenomena. Rather, its main role is to
provide a suitable common language for describing the effects. The development
of more detailed explanations and models of multitasking while driving, and other
behavioural effets, is an important area for further research.

4.5.3 Driver Behaviour and Accident Risk

Understanding the relation between behavioural effects and actual accident risk is
one of the most important, and difficult, issues in current traffic safety research.
The difficulties are due to a lack of sufficiently detailed behavioural data in existing
accident databases as well as the lack of appropriate behavioural models. A further
problem is that the usual interpretation of risks assumes both decomposability and
linearity. Yet both of these qualities are absent in dynamic, complex environments
such as traffic.
Based on the proposed framework, the relation between driving performance
and risk (for an individual driver) can be thought of as a complex function involving
(at least) the following factors:

1. The current complexity/difficulty of the driving task.

2. The driver's vehicle handling skills (performance on the tracking layer).
3. The ability to make a correct situation assessment (performance on the moni-
toring layer).
 4. A General Conceptual Framework for Modelling Behavioural Effects 79

4. The ability to adopt safety margins that are appropriate to (1) and (2), based on
(3) (determines performance on the regulation layer).
5. The effort spent on the control tasks on the different layers (2-4).

Based on this view, it is clear that performance degradation on one control layer
does not automatically increase accident risk. For example, reduced tracking per-
formance (e.g. due to visual distraction) is most risky if not properly compensated
for on the regulating layer (e.g. by reducing speed). Thus, risk (as a function of
performance) must be understood in terms of the relation between performance
on the different layers, where inadequate adaptation to the current driving condi-
tions and ones own driving skills could be hypothesised to be a critical factor. A
typical example of this is drunk driving, where alcohol is well known to induce
overestimation of the own performance driving capability. Thus, in this case, the
erroneous safety margin setting (on the regulation layer) is due to the cognitive
impairment induced by the drug. This line of reasoning also applies to individ-
ual differences with respect to risk taking. For example, the over-involvement of
young male drivers in accidents could be understood as in terms of overestimation
of driving performance combined with a higher propensity for sensation seeking
(i.e. different goals on the targeting level), leading to inadequate safety margin
settings. Yet another example is run-off-road accidents due to slippery roads. In
this case, the erroneous safety margin setting (reflected, e.g. in too high speed
in a curve) is due to an erroneous situation assessment on the monitoring layer,
which affects the regulatory level and finally induces instability and breakdown of
the tracking control. As discussed in the previous section, there is evidence that
cognitively loading tasks, such as phone conversation, impairs the ability to set ap-
propriate safety margins and adapt accordingly. However, the safety consequences
of this are still unknown.
It is very difficult to determine whether adequate adaptation has been achieved
in a particular situation, especially in terms of quantitative driving performance
metrics. One potential approach is to look for violations ofsafety margins. Possible
metrics of such violations include the amount of involuntary lane departures, min-
imum TLCITTC value, the amount of TLC/TTC values below some critical value
(van der Horst, 1990), or the total time spent below the critical value (time-exposed
TTC-TET; Minderhoud and Bovy, 2001). However, a basic problem with this
is that the accepted safety margins generally differ substantially between drivers
(and possibly also varies over time for an individual driver). Moreover, there is
yet no hard empirical data showing how these metrics relate to actual accidents.
Other key factors related to accident risk are expectancy and predictability (Victor,
2005). The ability to predict is central to remain in control and the occurrence of
unexpected events increases the risk for losing control. An issue of key impor-
tance is the ability of the driver to regain control. This is more difficult if the loss
of control also involves the loss of goals (e.g. on the regulating level).
While the COCOMIECOM framework is suitable for describing and reasoning
about these issues, it is clear that accident causation cannot be explained only in
terms of inadequate adaptation, or reduced predictability, but rather in terms of the
80 Engstrom and Hollnagel

interaction of multiple (behavioural, technological or organisational) precipitating

factors. This requires models that deal specifically with accident causation, e.g.
models based on Human Reliability Analysis (e.g. Hollnagel, 1993, 1998). How-
ever, the COCOMIECOM framework is a useful starting point for describing and
analysing the behavioural factors involved in accident causation.

4.6 Discussion and Conclusions

The objective of the present paper was to outline a conceptual framework for
describing behavioural effects of driver support functions. A key starting point
was the view of behaviour as a goal-directed, situated and dynamic activity. Driver
behaviour involves the simultaneous pursuit of multiple goals that may be more
or less related to driving itself. The COCOMIECOM hierarchical control model
(Hollnagel and Woods, 2005) was proposed as the basis for the framework.
It was shown that the framework offers a coherent taxonomy for categorising
a wide range of driver support functions with respect to their functional purpose
(rather than the underlying technology). This could be very useful, e.g. for clearly
defining the scopes of different development and evaluation methodologies and
standards (such as, e.g. the European Statement of Principles on Human Machine
Interface; Commission of the European Communities, 2000) in terms of the types
of functions that they apply to.
The second intended application area was the conceptualisation of behavioural
effects of driver support functions. It was demonstrated that the framework yields
coherent descriptions of a range of different effects, from behavioural adaptation
to the effects of multitasking. Moreover, it was argued that the COCOMIECOM
framework accounts for many aspects that are missed when a traditional informa-
tion processing/workload-based perspective is adopted, for example the radically
different effects of visual and cognitive secondary task load on driving perfor-
mance.
The third intended application area concerned the understanding of accident
causation. It was illustrated how the framework could provide a useful starting
point for addressing the difficult question of how driving behaviour/performance
is related to risk. A main conclusion was that accident risk cannot generally be
explained by a single aspect of driving performance. According to the proposed
framework, risk needs to be understood in terms of the relation between different
levels of performance. Thus, for example, a certain amount of weaving in the lane
may not be too risky as long as it is appropriately compensated for by slowing
down or increasing headway. It should also be noted that a lack of behavioural
change, e.g. a failure to reduce speed when entering a slippery road segment, may
be safety critical as well. Thus, inadequate adaptation can be proposed as one key
behavioural factor related to risk.
The present framework is based on the view of the human driver as an active
agent that, in most situations, is able to maintain control of the vehicle. A key
advantage of the framework is that it is well suited to describe adaptive behaviour,
which is a prevalent effect of all types of driver support functions. This contrasts
 4. A General Conceptual Framework for Modelling Behavioural Effects 81

to the traditional information processing models, which tend to view the human
as a passive receiver of information, subject to overload if the limited capacity is
exceeded. Another key property of the framework is the characterisation of driving
performance in terms of hierarchical control. The idea of hierarchical driver models
is certainly not new and the ECOM does not contradict existing models such as,
e.g. Alexander and Lunenfeld (1986) and Michon (1985). However, the added
value brought by ECOM is that it offers an account for how the different layers are
related, e.g. how a change in a high-level goal (such as realising that one is late to
the airport) can change performance on lower control layers (such as reducing the
accepted safety margins).
As stressed throughout this paper, COCOMIECOM are functional models that
make very few assumptions about internal cognitive structures and processes. For
present purposes, these models provide the starting point for a general conceptual
framework, based on which more specific models can be developed. Such models
should include both functional models as well as more detailed structural models
of the cognitive/neural mechanisms underlying behavioural effects of driver sup-
port functions (see, e.g. Victor, 2005 for a comprehensive review of models of the
latter type). Hence, functional and structural cognitive models should be viewed
as complementary and useful for different purposes. In addition to the further de-
velopment of more detailed driver models, the proposed conceptual framework is
a suitable starting point for framing hypotheses for empirical work on behavioural
effects of driving support functions. It is also useful as a common language in
industrial development of driver support functions, which helps maintaining the
focus on the purpose of the functions (e.g. how they should change driver be-
haviour), rather than their technological implementation (e.g. what type of sensors
are used).

References
Abelson, R.P. (1963). Computer simulation of "hot" cognition. In S.S. Tomkins & S Messick
(Eds.). Computer simulation ofpersonality. John Wiley and Sons, New York.
Alexander, G.J. and Lunenfeld, H. (1986). Driver expectancy in highway design and traffic
operations. US Department of Transportation. Report No: FHWA-TO-86-1.
Amditis, A., Polychronopoulos, A., Belotti, F. and Montanari, R. (2002). Strategy plan
definition for the management of the information flow through an HMI unit inside a car.
e-Safety Conference Proceedings, Lyon.
Bloch, H., Chemama, R., Gallo, A., Leconte, P., Le Ny, 1.-F. and Postel, 1. et aI., (Eds.)
(1999). Grand Dictionnaire de La Psychologie. Larousse, Paris.
Brookhuis, K.A., de Vries, G. and de Ward, D. (1991). The effects of mobile telephoning
on driving performance. Accident Analysis and Prevention, 23(4), 309-316.
Brostrom, R., Engstrom, 1., Agnvall, A. and Markkula, G. (2006). Towards the next gener-
ation intelligent driver information system (lDIS): The Volvo Cars Interaction Manager
concept. In Proceedings of the 2006 ITS World Congress, London.
Cacciabue, P.C. and Hollnagel, E. (2005). Modelling driving performance: A review of
criteria, variables and parameters. In L. Macchi, C. Re and P.C. Cacciabue (Eds.). Pro-
ceedings of the International Workshop on Modelling Driver Behaviour in Automotive
Environments. Ispra, Italy.
82 Engstrom and Hollnagel

Commission of the European Communities (2000). Annex to Commission Recommenda-

tion of 21 December 1999 on safe and efficient in-vehicle information and communication
systems: A European statement of principles on human machine interface. Official Jour-
nal ofthe European Communities L19, 25.1.2000, p. 64 ff, Brussels, Belgium: European
Union.
Damasio,A. (1994). Descarte'serror: Emotion, reason, and the human brain. G.P. Putnam's
and Sons, New York.
Dennett, D.C. (1987). The intentional stance. MIT Press, Cambridge, MA.
de Waard, D. (1996). The measurement ofdrivers' mental workload. ISBN 90-6807-308-7.
Traffic Research Centre. University of Groningen.
Dingus, T.A. (1995). Moving from measures of performance to measures of effectiveness
in the safety evaluation of ITS products or demonstrations. Paper presented at the safety
evaluation workshop. University of Iowa.
Donges, E. (1978). A two-level model of driver steering behaviour. Human Factors, 20(6),
691-707.
Engstrom, 1., Arfwidsson, 1., Amditis, A., Andreone, L., Bengler, K. and Cacciabue, P.C.
et aI., (2004). Meeting the challenges of future automotive HMI design: Overview of the
AIDE integrated project. In Proceedings, ITS in Europe. Budapest.
Engstrom, 1., Johansson, E. and Ostlund, 1. (2005a). Effects of visual and cognitive load in
real and simulated motorway driving. Transportation Research Part F, 8,97-120.
Engstrom, 1., Aberg, N., Johansson, E. and Hammarback, 1. (2005b). Comparison be-
tween visual and tactile signal detection tasks applied to the safety assessment of
in-vehicle information systems. In Proceedings of the Third International Driving Sym-
posium on Human Factors in Driver Assessment, Training and Vehicle Design. Rockport,
Maine.
Fosser, S., Saetermo, I.F. and Sagberg, F. (1997). An investigation of behavioural adaptation
to airbags and antilock brakes among taxi drivers. AccidentAnalyis and Prevention, 29(3).
Fuller, R.G.C. (1984). A conceptualisation of driving behaviour as threat avoidance. Er-
gonomics, 27, 1139-1155.
Fuller, R.C.G. (2005). Towards a general theory of driver behaviour. Accident Analysis and
Prevention, 37,461-472.
Gibson, J.1. (1979). The ecological approach to visual perception. Houghton Mifflin,
Boston.
Gibson, J.1. and Crooks, L.E. (1938). A theoretical field-analysis of automobile-driving.
The American Journal of Psychology, 51(3), 453-471.
Godthelp, H., Milgram, P.and Blaauw, J. (1984). The development of a time-related measure
to describe driving strategy. Human Factors, 26(3), 257-268.
Greenberg, 1., Tijerina, L., Curry, R., Artz, B., Cathey, L. and Grant, P. et al. (2003). Evalua-
tion of driver distraction using an event detection paradigm. Journal ofthe Transportation
Research Board, 1843. TRB, Washington DC.
Hale, A.R., Stoop, J. and Hommels, J. (1990). Human error models as predictors of accident
scenarios for designers in road transport systems. Ergonomics, 33, 1377-1388.
Harbluk, 1.L. and Noy, Y.I. (2002). The impact of cognitive distraction on visual behaviour
and vehicle control. Report no. TP 13889E. Transport Canada, Ontario.
Hollnagel, E. (1993). Human reliability analysis: Context and control. Academic Press,
London.
Hollnagel, E. (1998). Cognitive reliability and error analysis method. Elsevier Science,
Oxford, UK.
Hollnagel, E. (2004). Barriers and accident prevention. Ashgate Publishing Limited, Alder-
shot.
 4. A General Conceptual Framework for Modelling Behavioural Effects 83

Hollnagel, E. (2006). Outline of a function-centred approach to joint driver-vehicle system

design. Cognition, Technology & Work (in print).
Hollnagel, E., Nabo, A. and Lau, I. (2003). A systemic model for driver-in-control. Pro-
ceedings of the 2nd International Driving Symposium on Human Factors in Driver As-
sessment, Training, and Vehicle Design., Park City, UT. Public Policy Center, University
of Iowa.
Hollnagel, E. and Woods, D.D. (2005). Joint cognitive systems: Foundations of cognitive
systems engineering. Taylor and Francis/CRC Press, Boca Raton, FL.
Horrey, W.J. and Wickens, C.D. (2004). The impact of cell phone conversations on driv-
ing: A meta-analytic approach. Technical Report AHFD-04-2/GM-04-1, General Motors
Cooperation, Warren, MI.
Jamson, A.H. and Merat, N. (2005). Surrogate in-vehicle information systems and driver
behaviour: Effects of visual and cognitive load in simulated rural driving. Transportation
Research Part F, 8, 79-96.
Kahneman, D. (1973). Attention and effort. Prentice Hall, Englewood Cliffs, N1.
Lee, D.N. (1976). A theory of visual control of braking based on information about time-
to-collision. Perception, 5, 437-459.
Marken, R.S. (1986). Perceptual organisation of behaviour: A hierarchical control model
of coordinated action. Journal ofExperimental Psychology: Human Perception and Per-
formance, 12,267-276.
Markkula, G. and Engstrom, 1. (2006). A steering wheel reversal rate metric for assessing
effects ofvisual and cognitive secondary task load. In Proceedings ofthe 2006 ITS World
Congress, London.
McRuer, D.T., Allen, R.W., Weir, D.H. and Klein, R.H. (1977). New results in driver steering
control models. Human Factors, 19,381-397.
Michon, J.A. (1985). A critical review of driver behaviour models: What do we know? What
should we do? In Evans L.A. and Schwing R.C. (Eds.). Human Behaviour and Traffic
Safety. Plenum Press, NY, pp. 487-525.
Minderhoud, M.M. and Bovy, P.H.L. (2001). Extended time-to-collision measures for road
traffic safety assessment. Accident Analysis and Prevention, 33, 89-97.
Moray, N. (1967). Where is attention limited? A survey and a model. Acta Psychologica,
27,84-92.
OECD (1990). Behavioural adaptations to changes in the road transport system. Report
Prepared by an OECD Expert Group. Road Transport Research Programme.
Neale, V.L.,Dingus, T.A., Klauer, S.G., Sudweeks, 1. and Goodman, M. (2005). An overview
of the 1DO-car naturalistic study and findings. Paper presented at the 19th International
Technical Conference on Enhanced Safety of Vehicles (ESV). Washington DC, June 6-9.
Neisser, U. (1976). Cognition and reality. Freeman, San Francisco.
Nilsson, R. (2001). Safety margins in the driver. PhD Thesis. Uppsala University, Sweden.
Ostlund, 1., Nilsson, L., Carsten, 0., Merat, N., Jamson, H. and Jamson, S. et aI., (2004).
Deliverable 2 - HMI and Safety-Related Driver Performance (No. GRD1/2000/25361
S12.319626). Human Machine Interface and the Safety of Traffic in Europe (HASTE)
Project.
Powers, W.T. (1998). Making sense ofbehavior-the meaning ofcontrol. Benchmark, New
Canaan, CT.
Ranney, T. (1994). Models of driving behaviour. A review of their evolution. Accident
Analysis and Prevention, 26(6), 733-750.
Rasmussen,1. (1983). Skills, rules, and knowledge: Signals, signs, and symbols, and other
distinctions in human performance models. IEEE Transactions on Systems, Man, and
Cybernetics, SMC 13, 257-266.
84 Engstrom and Hollnagel

Recarte, M.A.and Nunes, L.M. (2003). Mental workload while driving: Effects on visual
search, discrimination, and decision making. Journal of Experimental Psychology, 9.
Saad, F., Hjalmdahl, M., Canas, 1., Alonso, M., Garayo, P. and Macchi, L. et aI., (2004).
Literature review of behavioural effects. Deliverable 1.2.1, AIDE Integrated Project,
Sub-project 1. IST-I-507674-IP.
SAVE-IT (2002). Safety vehicle( s) using adaptive interface technology (SAVE-IT) Program,
DTRS57-02-20003. US DOT, RSPSNolpe National Transportation Systems Center (Pub-
lic Release of Project Proposal), www.volpe.dot.gov/opsad/saveit/index.htmI.
Salvucci, D.D. (2001). Predicting the effects of in-car interface use on driver performance:
An integrated model approach. International Journal of Human-Computer Studies, 55,
85-107.
Shinar, D. (1993). Traffic safety and individual differences in drivers' attention and infor-
mation processing capacity. Alcohol, Drugs and Driving, 9, 219-237.
Smiley, A. (2000). Behavioural adaptation. Safety and Transportation Systems Research
Record, 1724, 47-51.
Simon, H.A. (1955). A behavioural model of rational choice. The Quarterly Journal of
Economics, LXIX, 99-118.
Summala, H. (1985). Modeling driver behavior: A pessimistic prediction? In Evans L,A"
Schwing R.C. (Eds.). Human Behaviour and Traffic Safety. Plenum Press, New York,
pp.487-525.
Summala, H. (1988). Risk control is not risk adjustment: The zero-risk theory of driver
behaviour and its implications. Ergonomics, 31(4),491-506.
van der Horst, R. (1990). A time based analysis of road user behaviour in normal and
critical encounters. PhD Thesis. Delft University of Technology, The Netherlands.
van der Horst, R., Godthelp, H. (1989). Measuring road user behaviour with an instrumented
car and an outside-the-vehicle video observation technique. Transportation Research
Record # 1213, Washington DC: Transportation Research Board, 72-81.
Weir, D.H., McRuer, T.M. (1968). A theory of driver steering control of motor vehicles.
Highway Research Record, 247, 7-39.
Wickens, C.D. (2002). Multiple resources and performance prediction. Theoretical Issues
in Ergonomics Science, 3(2), 159-177.
Victor, T.W. (2005). Keeping eye and mind on the road. PhD Thesis. Uppsala University,
Sweden.
Victor, T.W., Harbluk, 1.L. and Engstrom, 1. (2005). Sensitivity of eye-movement measures
to in-vehicle task difficulty. Transportation Research Part F, 8, 167-190.
Wilde, GJ.S. (1982). The theory of risk homeostasis: Implications for traffic safety and
health. Risk Analysis, 2, 209-225.
5
Modelling the Driver in Control
BJORN PETERS AND LENA NILSSON

5.1 Introduction

Modelling driver behaviour can be done with different purposes. One obvious aim
could be to provide a model of what drivers actually do and explain observed be-
haviour with or without support systems. Other objectives can be to discriminate
between safe and unsafe driving. Furthermore, a model can be used to imple-
ment a dynamic real-time control of driver support functions (see Chapter 6). The
following text is written in line with the first objective.
Safe driving requires that the driver is in control of the vehicle and the driving
situation. Driver support systems aim to help the driver and facilitate driving in
one way or the other. Support systems should be a response to identified existing
or potential problems related to the driving task. New technology like ADAS and
IVIS can facilitate driving but it could, on the other hand, make driving even
more complex and demanding. Solving one problem could be done at the cost of
introducing another. Thus, we need to understand driving task demands, driver
behaviour and how support systems can influence the driver's control and safety.
Here a broad overview of different theoretical approaches is given with focus
on control and safety. The use of a cognitive systems engineering approach is
advocated to investigate pros and cons of driver support systems. In the end, an
example is given on how findings from an experiment with a joystick controlled
car can be interpreted.

5.2 A Cognitive View of Driving

Driving is one of the most complex and safety critical everyday tasks in modern
society (Groeger, 2000). Driving a car is complex in the sense that it requires
the driver to employ a wide range of abilities in order to interact with a complex
environment and to manage the driving task demands. Driving is dynamic as the
demands can change back and forth from very low to extremely high, sometimes
within fractions of a second. When demands are high, driving is carried out in a

85
86 Peters, VTI and Nilsson, VTI

force-paced fashion; while as the demands are low, it can be performed in a more
self-paced manner. Normal driving can be considered as a cognitively motivated
and controlled task. Cognition is here used in the pragmatic sense as defined by
Neisser (1976), that is, cognition in context. A motive for this cognitive stance
can be found in Michon's (1985) visionary talk on driver behaviour modelling
' . .. the distinctly hierarchical cognitive structure of human behaviour in the traffic
environment . . . '. However, a cognitive approach does not imply that the driver's
perceptual and psychomotor abilities are to be neglected. Thus, driving can be
viewed as a cognitive task of control in a context perceived through the senses and
manipulated with control actions based on unconscious (automated) or conscious
decisions. This cognitive approach has been applied , for example , in adaptive
control models, control theory and cognitive systems engineering in order to model
driver behaviour.

5.3 Human Abilities

The human controller can be described in functional terms . In order to carry out
the driving task, the driver utilises three, tentatively different, functional abilities:
Cognitive, perceptual and motor abilities (see Fig. 5.1) . Cognitive abilities include,
for example, memory, decision, attention and supervision. Perceptual abilities can
be, for example, visual, auditory, tactile and proprioceptive. Finally, motor abilities
relate to physical dimension, motion and force , for example, reach, force and
endurance. These human abilities can be developed into very efficient functions
by train ing and experience. The abilities should not be considered as separate
entities but as highly interactive and even more so as they develop into skilled
behaviour. This functional description of the driver can be utilised to depict the
driver from a cognitive systems engineering view.

Cognitive
* mem OlY

21'
* attent on
Perceptual * decision
* vision
* hearing
* touch ~
* proprioception ----v
Physical
* size
* reach
*force ~
* endurance ~
FIGURE 5.1. Humanabilities em-
ployed by the driver.
 5. Modelling the Driver in Control 87

TABLE 5.1. Classification of driver behaviour models (adapted

from Michon, 1985).
Taxonomic Functional
Behavioural (input-output) Taskanalyses Adaptive control models
Internal state (psychological) Trait models Motivational models

5.4 Classifying Driver Behaviour Models

Several theories and models of driver behaviour have been presented, applied, anal-
ysed, criticised and abandoned. So far, no all-purpose, generic, comprehensive and
verifiable model of driver behaviour has been presented (Ranney, 1994). Differ-
ent models often emphasise specific and sometimes different aspects of driver
behaviour, for example, accident causation, education and training, behavioural
adaptation. Instead of searching for the ultimate model, it seems more feasible to
use a generic framework in order to get a structured view of existing driver be-
haviour models. Two different modelling approaches can be distinguished. The first
is to model what drivers actually do when driving. The second is to describe what
the drivers should do by modelling the driving task itself. The first can be called a
behavioural approach and the second a normative approach. Michon (1985) pro-
posed a generic classification of driver behaviour models, using a two-dimensional
classification (see Table 5.1). Firstly, he distinguished between behavioural or
input-output oriented models and internal state or psychological models. Sec-
ondly, he differentiated between functional and taxonomic models. Taxonomic
models are inventories of isolated facts, while functional models specify compo-
nents of driver behaviour and their dynamic relations. In this way he distinguished
between four types of models: task analysis models, trait models, adaptive control
models and motivational models. Task analysis models decompose driving into
tasks and subtasks and relate them to driver requirements and abilities. Trait mod-
els are based on the idea that it is possible to identify the accident-prone driver
with the use of well-designed tests. Functional models differ from taxonomic in
that they connect model components in order to consider the dynamics of driver
behaviour, for example, hierarchical structures. Adaptive control models apply
functional approach to capture behavioural changes. Finally, motivational models
consider internal states as attitudes, subjective risk and insight as controlling fac-
tors. The models that will be discussed in the following can be mostly categorised
as adaptive control models. For a discussion on other types of models, see Carsten
(2005, Chapter 6).

5.5 Hierarchical Control Models

Most models can offer only post hoc explanations of observed behaviour
or possibly explain aggregated accident data (Rumar, 1988). Early cognitive
models focused on information processing, and driving was considered as a
88 Peters, VTI and Nilsson, VTI

problem-solving task. Accidents were attributed to incorrect information process-

ing, and Rumar concluded that there was an urgent need to develop models and
hypothesis that can predict actual driver behaviour. Michon (1985) claimed that
the lack of progress emerged from the failure to consider results from cognitive
psychology. Later models incorporate developments from cognitive psychology,
for example, hierarchical control structures and automaticity (Ranney, 1994).
Rumar (1988) pointed out that an important behavioural uniqueness of driving
can be its combination of consciously controlled (cognitive) and unconsciously,
automatic (perceptual) behaviour. Hierarchical control models became more
accepted when Michon (1985) advocated the idea that such an approach would
resolve some of the identified shortcomings with earlier models. However, this
was not a new idea, for example, Allen et al. (1971), McRuer et al. (1977) and
Janssen (1979) described driving as a hierarchical structured task with strategic,
tactical and operational components demanding different levels of driver control
much earlier. At the strategic level, the driver is concerned with tasks such as
planning the journey, selecting the mode of transport and choosing a route. At the
manoeuvring level, the tasks concerned include overtaking, giving way to other
vehicles and obeying traffic rules. At the control level, the driver is concerned
with controlling the vehicle, for example, controlling speed, following the road
and quite simply keeping the car on the road. The model assumes an interaction
between the three levels in which goals and criteria are defined at a higher level
and the outcome of lower levels modifies goals at a higher level. Finally, Michon
(1985) meant that a comprehensive model of driver behaviour should not just
identify different levels of control but also provide a control structure that enables
control to shift from one level to another in a timely manner.
The hierarchical structure of the driving task can be matched to actual driver
behaviour. Human control structures are highly flexible and highly dependent on
practice and experience. These structural aspects of human performance were ad-
dressed in the hierarchical skill-, rule- and knowledge-based behaviour (SRK)
model developed by Rasmussen (1986). Rasmussen identified a number of cog-
nitive functional elements organised in a three-layered structure. The model dis-
criminates between skill-based, rule-based and knowledge-based controls. The
main issue in Rasmussen's framework is the hierarchical nature of human control
replacing the serial model used in early models of human information process-
ing. The SRK model has been extensively used to model driver behaviour. The
hierarchical control model takes into account both the hierarchical structure of the
driving task and the driver behaviour by combing the two frameworks mentioned
above (Ranney, 1994). Thus, Michon (1985) compiled a two-dimensional matrix
that has been used to explain driver behaviour and to identify driver support needs
(Michon, 1993; Ranney, 1994; Nilsson et aI., 2001) with driving task demand and
driver behaviour as the two dimensions (see Fig. 5.2). New driver support systems
might force the driver to learn new skills and forget old skills. As a consequence,
the driver might have to shift from skill-based behaviour to knowledge-based be-
haviour in order change to new skill-based behaviour, for example, learning to use
an ADAS type of system like ABS.
 5. Modelling the Driver in Control 89

Skill based

Strategical

Driving
Task Tactical
Demands I - - - -I -- - - -+-____

Control

FiGURE 5.2. IVIS and ADAS can influence both driving task demands and driver behaviour
(from Nilsson et al., 2001) .

With the hierarchical structuring of both task and behaviour, it becomes evident
that time is an aspect of driving that should be considered. Time constraints are im-
plicitly different for the three task levels, even if time is not specifically addressed
in the hierarchical control model. The following approximate time frames apply:
lOs or more for tasks at strategic level, between I and lOs at tactical level and less
than I s at control level. Finally, it should be noted that more complicated driver
tasks, for example, performing route planning while driving or mobile phoning
during driving may require knowledge-, rule- and skill-based actions in combina-
tion . It is not simply the time requirements that distinguish the different levels of
tasks, but also the requirements of attention, workload and the consequences of
mistakes, etc. The hierarchical control models are functional models, which do not
specifically include motivational aspects and should be classified as adaptive con-
trol models . Furthermore, they do not explicitly consider the context, for example,
controller's influence on the system to be controlled (Hollnagel, 2000) .

5.6 Control Theory

Control theory or cybernetics is a general theory aimed at understanding self-
regulating systems (Wiener, 1954; Ashby, 1956; Carver and Schreier, 1982). The
systems view approach applied in cybernetics distinguishes itself from the more
traditional analytic approach by emphasising the interactions and connectedness
of the different components of a system . The basic control theory idea is simply
based on a stimulus-response loop . The system current status is compared to a
reference value and the deviation between system current status and reference
determine the control action. However, the definitions of system boundaries and
the relation between cause and effect are different from, for example, information-
processing models . Control theory has been used to describe and study a wide
range of systems from ' pure' technical to biological and social systems . Adaptive
control models are also based on control theory.
90 Peters, VTI and Nilsson, VTI

Stimulus

Environment

-------------: Response/Stimulusr -------------

Action I _;Sensation
--[----------: EnvIronment:---------- ----
I--------------'"j I---------------J
1 1 1 I

1
I
Decision ~
1 1
Perception .I
~ J ~-------------..!

FIGURE 5.3. Open loop - closed system versus closed loop - open system adapted from
Jagacinski and Flach (2003).

A system is an abstract construct, used to identify the focus of interest (Jagacin-

ski and Flach, 2003). The system is also sometimes contrasted to the environment.
A system typically refers to the phenomenon of interest and the environment
to everything else. When there is a sharp boundary between the system and the
environment, the system is considered as a closed system. The environment does
not influence the system and can be disregarded. In a closed system, control is
well defined and deterministic. Such a closed systems view is applied in human
information-processing (HIP) models (e.g., Wickens, 1992) when the controller
(e.g. driver) is considered as a closed system and the controlled object (e.g. car)
belongs to the environment. The rationale for this approach was to divide human
behaviour into isolated entities (e.g., sensation, perception, decision and action)
which could be studied independently (see Fig. 5.3, top). As can be seen, it is a
closed system (detached from the environment) and the control is an open-loop;
that is, there is no feedback depicted from the controller's output to system input.
However, the demarcation between a system and the environment is often not
that sharp. Or rather there are phenomena that cannot be explained with a closed
systems view. This is specifically true when concerned with behavioural science
where the relationship between, for example, a controller and the environment
seems to be of prime importance in order to understand complex phenomena like
behavioural adaptation. When the demarcation between a system and the environ-
ment is diffuse, the system can be considered an open system. All socio-technical
systems can be viewed as open systems (Flach, 1999). The primary difference with
other SR models is that control theory considers the operator's influence on the
controlled object and the system boundaries are defined so that both the operator
and the controlled object are included. This open system view is more consistent
with an ecological view (Gibson, 1966). System boundaries can be very tight to
all embracing depending on the focus. This open system view can be depicted by
closing the loop and position the environment in the centre (see Fig. 5.3, bottom).
 5. Modelling the Driver in Control 91

By closing the loop, the cause - effect relation between stimulus and response
loses its meaning. In a closed-loop system the stimulus and response are tightly
linked and there is no clear distinction between what is cause and what is effect.
This restructuring was done to illustrate that the stimuli are as much determined
by the actions as the actions are determined by the stimuli (Jagacinski and Flach,
2003). In this way we have a closed-loop open system, as the environment is
included and the interdependencies between the controller and the environment
are considered. This view can be useful when modelling driver behaviour and, in
particular, driver behaviour adaptation and driver support systems.
Cognitive Systems Engineering (CSE) put cognition in context by applying an
overall systems view. CSE decomposes complex tasks along two dimensions ab-
stract - concrete and whole - part (Jagacinski and Flach, 2003). In this way the
focus is shifted from internal functions of either humans or machines to the ex-
ternal function making up a joint cognitive system including both the controller
and the controlled system (Hollnagel and Woods, 2005). The driver will be viewed
as a controlling system and the car as a technical system to be controlled. In this
perspective the car will constitute the primary context for the driver. As more of the
context is included, the system will expand as a multi -layered functional descrip-
tion. The system boundaries have to be defined according to the purpose of the
analysis. The two most important concepts in Control Theory are the open/closed
system and the open/closed loop control. Jagacinski and Flach (2003), among oth-
ers (e.g. Hollnagel and Woods, 2005; Hollnage12002), have further developed the
concepts of control theory. Jagacinski and Flach have also provided some quan-
titative tools based on control theory that can be used to capture and understand
also qualitative aspects of human performance, for example, driving behaviour and
driver support systems.

5.7 Adaptive Control Models

Adaptive control models are concerned with issues of how the driver adapts his or
her control to the characteristics of the system to be controlled (driver-vehicle-
environment). Two categories of adaptive control models can be distinguished
servo-control models addressing continuous tracking and information flow control
models addressing discrete decision making (Michon, 1985). However, both views
can be used in a complementary manner to understand some aspects of manual
control. For example, firing the motor neurons is a set of discrete all-or-none
response. However, the motion of an arm depends on the integration over many
neurons and can best be described as continuous at a coarse time scale. Early
servo-control models were mostly used to study compensatory (feedback) steering
control performance on roads with varying curvature and for evasive manoeuvres.
McRuer et al. (1977) questioned this way of modelling steering behaviour, as it
did not consider the anticipatory control that seems to be a behaviour needed in
order to understand skilled driving behaviour.
92 Peters, VTI and Nilsson, VTI

External _ _-----,
disturbances
Lateral
position

Heading
angle

Desired
path

FIGURE 5.4. A three-level servo-control model of steering (from McRuer et al., 1977).

McRuer et al. (1977) proposed a three-level servo-control model of drivers'

steering behaviour. First of all, they described driving as consisting of a hierarchy
of navigation, guidance and control phases conducted simultaneously with visual
search, recognition and monitoring operations. They also distinguished between
closed-loop (compensatory) control and open-loop (anticipatory) control. Com-
pensatory steering was described as two feedback loops (see Fig. 5.4). Firstly,
the lateral position is fed back and compared to the desired path, and if there
is a deviation it will result in an error-correcting action, which is compared to
current heading angle and, if needed, a steering wheel correction will be made.
The perceived road curvature derived from visual input guides the pursuit con-
trol. Secondly, pursuit control is an open-loop feed-forward control element that
permits the driver to follow the anticipated road curvature. An interesting third
concept is the precognitive control that in practice is a first phase of dual-mode
control, that is, both open- and closed-loop controls. Precognitive control consists
of previously learned control actions, which are triggered by situation and vehicle
motion but work as pure open-loop control.
In view of this model, steering can be considered in terms of output as a position,
velocity or acceleration control system (Jagacinski and Flach, 2003). If tyre angle
is considered to be the output, then the steering system can be approximated as a
position control system with a gain given by the steering linkage. While, if heading
angle is viewed as the output, then steering can be considered a rate control system.
In this case the gain is proportional to the velocity of the front wheels. Finally,
if lateral position is considered the output, then steering should be viewed as an
acceleration control system. If so, then the effective gain between steering wheel
angle and lateral position is proportional to the square of the velocity. This shows
that lateral and longitudinal control of the car is not independent but very much
entwined. For a further discussion on control theory modelling, see also Wier and
Chao (2005, Chapters 16 and 17).
Even if McRuer and his colleagues (1977) described driving as a hierarchical
task, they concentrated very much on steering control. They largely considered
driving as a closed system and disregarded the environment except road geometry.
Michon (1985) cited Reid (1983), who concluded that the servo-control model
 5. Modelling the Driver in Control 93

described above cannot successfully cope with driver tasks other than follow-
ing straight and smoothly curved roads. The model needs to be better integrated
with the guiding visual environment as described by, for example, Gibson (1966).
Michon concluded that 'The two fields - perception and vehicle control - are
still lacking a theoretical integration. Combining them would constitute a major
breakthrough, ... ' .

5.8 Cognition in Control

Most of the models discussed so far are basically mechanistic, as they do not
recognise the need of higher order cognitive abilities. The interaction between the
driver and the environment is not explicitly included in the model but rather seems
to be an implicit presumption. A useful model needs to be better connected with
the context. The lack of context in cognition was addressed by Neisser. Neisser
(1976) criticised the concept of direct perception (Gibson, 1966) by stating that
to see is not just to perceive but also to interpret and understand on a conscious
level. Neisser (1976) proposed a cognitively driven model of perception called
the perceptual cycle, which includes the interaction between the observer and the
environment. He introduced what he called anticipatory schemata that prepare and
control our perception. Neisser (1976) further meant that human control works in a
way similar to the perceptual cycle. Thus, to control a system, the controller has to
have a model of the system to be controlled. The importance of a control-guiding
model can also be understood in the light of 'The law of requisite variety' (Ashby,
1956), which states in principle that the variety of the controller should match the
variety of the system to be controlled. Thus, the controller's understanding of the
system that is being controlled will determine the actual control actions. In other
words the driver's mental model of the vehicle, other drivers, road condition, etc.,
will determine the driver's control behaviour.
Hollnagel and Woods (2005) described a cyclical model of control, the basis
of on the principles of Neisser's perceptual cycle. Hollnagel's cyclical control
model was used as the basis for the Contextual Control Model, which describes in
general terms how performance depends on perceiving feedback events, interpret-
ing and modification of current understanding, selection and execution of actions.
Driver control behaviour can be described as shown in Fig. 5.5, which is based on
Hollnagel's cyclical model of control (Chapter 4).
The control cycle is divided into three phases: perception, decision and action.
The control cycle is cognitively initiated by the driver depicted with the arrow
coming out of the driver's head. The driver's mental model of the system to be
controlled and the environment will guide the search for information during the
perceptual phase. The perceived situation is compared to a reference value defined
by the driver. The comparison is followed by a cognitive phase. During this phase
a decision will be made on the basis of the difference between reference value and
current situation, that is, the error. The aim will usually be to minimise the error.
This cognitive phase is followed by an action phase during which an appropriate
94 Peters, VII and Nilsson, VTI

FIGURE 5.5. The cognitive control

cycle adapted from Hollnagel' s
(Hollnagel and Woods, 2005)
d Externa l events

r'ITeA~
contextual control model.

action is selected and carried out. This action influences the environment depicted
by the outward arrow. Once the action is carried out, the driver searches and per-
ceives the effect of the action together with possible external events and the circle
is closed. The result of the action phase is also fed back to the driver and will
change the driver's mental model of the control loop and the system under con-
trol. In other words it is previous experience and outcome of the driver's actions
on the controlled system that will form and develop his or her mental models.
The three phases are described as three distinct entitie s but in reality the phases
might be overlapping and not separated as might appear from the figure. How-
ever, in principle the three phases are different in character. This model of driver
control provide s a foundation to capture the dynamics in driving, for example ,
compensatory closed-loop and anticipatory open-loop driving .
In relation to the discussed control model it can be interesting to consider the
dual control problem: The driver should both determin e the system status and at
the same time control it (Jagacinski and Flach, 2003). The driver occasionally has
to get out of control in order to maintain control. This was described by Weinberg
and Weinberg (1979) as the fundamental regulator paradox: The lesson is easiest
to see in terms ofan expe rience common to anyon e who has ever driven on an icy
road. The driver is trying to keep the carfrom skidding. Toknow how much steering
is required, she must have some inkling of the road' s slickness. But ifshe succeeds
in completely preventing skids, she has no idea how slippery the road really is.
Good drivers. expe rienced Oil icy roads, will intentionally test the steering from
time to time by 'jiggling ' to cause a small amount of skidding. By this technique
they intentionally sacrifice the perfect regulation they know they cannot attain in
any case. III return. they receive information that will enable them to do a more
reliable. though less pe rfect job. Unintentional skidding will, of course, provide
sufficient information to the driver on how to adapt the driving behaviour in order
to overcome the slippery road condition.
 5. Modelling the Driver in Control 95

5.9 Goals for Control

The control theory model discussed above assumes that there is a goal or a reference
value, which is used to determine appropriate actions. Different goals for safe
driving behaviour have been proposed, for example, zero risk, threat avoidance,
safety margins, response to risk (Naatanen and Summala, 1974; Fuller, 1984;
Summala, 1985, 1988; Groeger, 2000). However, one of the first models aiming
to describe safe (normal) driving behaviour was presented by Gibson and Crooks
(1938). They meant that driving is a task that is carried out in time and space and
described driving as a task of controlling the car within the field of safe travel.
They defined the field of safe travel as 'an indefinite bounded field consisting, at
any moment, of the field of possible paths which the car may take unimpeded' . It is
an imaginary dynamic area in front of the vehicle with a shape of an outstretched
tongue (see Fig. 5.6). Obstacles in the terrain mainly determine the boundaries
in the field. Gibson and Crooks' model provides a foundation for the concept of
safety margins and a guiding mechanism for normal safe driving behaviour.
Driving within safety margins is a concept that has been explored by several
researchers. As an example, the concept of the longitudinal safety margins was
explored by van der Hulst (1999). Van der Hulst meant that experienced drivers
have a mental model of the driving task and expectations about what will occur
in which situation. As a consequence, the driver can effectively scan the traffic
environment. The mental model and the expectations will allow the driver to build
up and preserve situation awareness and also to take advantage of a precognitive
control as described by McRuer et al. (1977). Furthermore, driving is a task that
can be carried out in many ways and will provide opportunities for behaviour

I
,I
I
I
I

,/
,./
,,/'.
.'

FIGURE 5.6. The concept of field of safe travel (from Gibson and Crooks [1938] published
with permission of The American Journal of Psychology).
96 Peters, VTI and Nilsson, VTI

adaptation. The driver can choose from several driving strategies. This is a view that
can help to better understand the mechanism behind the anticipatory behaviour. Van
der Hulst (1999) also meant that driving is a task that allows for pace adjustments
by means of adjustments in speed and other safety margin. The driver's choice
of speed and safety margins will determine the time available to react to relevant
changes in the environment. Van der Hulst thereby connected Gibson and Crooks's
'field of safe travel' with the Ashby's 'law of requisite variety'. Summala (1985,
1988) proposed that safety margins could be operationally defined as distance-
or time-related measures like Time-to-line-crossing (TLC) and Time-to-collision
(TTC). The concept of safety margins can also be used to explain, at least partly,
accident causation. In-depth accident studies have shown that late detection is a
very common explanation given for collisions (Rumar, 1988). Late detection can
be described as violation of safety margins. Thus, time is critical for safe driving,
which will be discussed later.
However, driver behaviour is most likely determined by a set of concurrent
goals. The goals might also shift during a drive. Given the hierarchical structure
of the driving task and driver behaviour as described above, it seems likely that
the driver applies different goals for different levels of control. Thus, the driver
can at the same time drive to reach a destination in time, stick to the traffic rules
and avoid accidents. Goals for driving behaviour can also differ between drivers
and situations. Furthermore, goals for driving can be extended to incorporate, for
example, goals for life, skill for living, sensation seeking, pleasure, etc. (Hatakka
et aI., 2002). Thus, it seems likely that the driver has to find a balance between
different goals. The idea of balancing between different goals was applied by
Wilde (1982) in the risk homeostasis theory. Homeostasis is originally a term
used to describe a complex mechanism for maintaining metabolic equilibrium in
biological systems. Wilde meant that driver behaviour is guided by a target risk
level that is determined by a combination of subjective and objective risk. An
implication of Wilde's approach is that all actions taken to improve safety will
be neutralised by the driver - at least on an aggregated level. The consequences
of Wilde's theory led to considerable controversy and the theory has been re-
jected by several researchers (e.g. McKenna, 1982; Michon, 1985; Sanders and
McCormick, 1993). According to Fuller (2005, Paper 10), Wilde based his the-
ory on a misinterpretation of empirical findings made by Taylor (1964). Fuller
proposed instead that task difficulty homeostasis, which describes the dynamic
interaction between driving task demands and driver capability, could be used
as key-sub goals to describe driver behaviour. Fuller argues that the task diffi-
culty homeostasis overcomes some problems with the safety margin concept and
conforms to the hierarchical structuring of the driving task and driver behaviour
and risk homeostasis theory can be viewed as a special case. Other human be-
haviour researchers has shown how homeostasis can be used to understand how,
for example, emotions can influence behaviour (Damasio, 1999; see Vaa, 2005,
Paper 12). Thus, it seems likely that homeostasis is a mechanism that could be
explored further to understand how different goals interact and determine driver
behaviour.
 5. Modelling the Driver in Control 97

5.10 Time and Time Again

Time considerations are crucial for the cognitive control cycle described above.
Hollnagel and Woods (2005) divided the control cycle in three different phases:
perception, decision and action. Time constraints can be incorporated into the
cognitive control cycle (see Fig. 5.7), slightly adapted from Hollnagel. Thus, speed,
road geometry, obstacles in the field of travel, sight conditions among a range of
other factors determine the time available for the control cycle. This time is labelled
Tu (usable time). The times needed to carry out the three phases (perceive, decide
and act) are labelled Tp , Td and Ta , respectively. In 'normal' driving Tp + Td + Ta
is less than Tu .
Depending on the situation, normal driving constitutes a combination of com-
pensatory and anticipatory control. The driver usually strives to balance between
compensatory and anticipatory driving. Anticipatory driving requires Tu to be
longer than (Tp + Td + Ta ) but as the time required comes close to what is avail-
able driving becomes more compensatory. If the total usable time is not sufficient,
performance will start to degrade. The control becomes more erroneous or slug-
gish and oscillatory (Jagacinski, 1977). Reducing speed is one way to gain time
and control. The model also depicts that the reason for deteriorated performance
can be attributed to prolonged perception, prolonged decision, prolonged action
or some combination of the three. In any case, the result will be that the used
time will be more than the usable time. All three phases are connected, meaning
that if one part requires less time than expected then there will be more time for
the remaining two phases. When traffic demands are low and the driver is ex-
perienced, evaluation, selection and even action require little time and there will
be plenty of time available. Time pressure is in this view a critical component
of driving behaviour when determining the driver's safety margins. Closely re-
lated to this is the concept of uncertainty. The driver will try to keep uncertainty
at an acceptable level to maintain pace control and safety margins (Godthelp,

external
events

Tu= total usable time

Tp=time needed to perceive
Td=time needed to decide
Ta=time needed to act

FIGURE 5.7. Time and control in the cyclic control model (adapted from Hollnagel, 2002).
98 Peters, VTI and Nilsson, VTI

Milgram and Blauw, 1984; Wierwille, 1993). The driver can apply an intermittent
sampling strategy to cope with lack of time for the control loop. However, this
strategy can, if maintained, increase the level of uncertainty, depending on the
situation (Lee, 2006). Increased uncertainty and lack of anticipation will make
drivers vulnerable to accidents. The time concept in the control cycle described
above can also be applied to driver support systems. If the driver's mental model
of a support system is incorrect or incomplete, this misunderstanding can pro-
long the time needed to complete the control loop. If the support system provides
delayed or contradictory feedback and the driver is not guided to the right ac-
tion, then the driver might eventually lose control. Lost control can also be due
to a situation when the driver requires a long time to perceive what the system is
doing.

5.11 Multiple Layers of Control

The driving task can be structured in hierarchical levels as discussed earlier. So
far we have mostly been concerned with low levels of control. To get a more
holistic view of driver behaviour, we need to expand the cognitive control loop to
incorporate more of the context and apply a joint systems view (driver-vehicle-
road-traffic etc.). The extended control model (ECOM) was developed by Holl-
nagel and Woods (2005) to consider aspects of joint system control. Hollnagel
distinguished four hierarchical layers of control: tracking, regulating, monitor-
ing and targeting, with targeting as top level and controlling at the bottom level.
These four layers correspond to the levels distinguished in a generic decision
model. Hollnagel pointed out that there is no absolute reference that can be used
to determine the number of layers needed for all cases. Rather, the purpose de-
termines the number of layers needed; for example, Powers (1998) identified 11
levels in his perceptual control theory. The layers in Hollnagel's model should
not be considered as distinctive but more as continuous and overlapping. The
control layers are connected in such a way that goals for control at lower layers
are determined at higher layers and feedback is provided from lower to higher
layers. The primary control of the car, that is, lateral and longitudinal motion of
the vehicle, involves all control layers even if tracking control is the most obvi-
ous. Furthermore, the driver has to perform several tasks - not just the primary
control of the car but also several additional tasks that are often carried out in par-
allel. Such additional tasks can be interaction with secondary car controls, driver
support systems and even nomad systems, for example, mobile phones, hand-
held computers, etc. Even if a driver support system can be aimed at a specific
control layer, it might very well affect control on other layers. For a more elabo-
rated description and application of the ECOM in relation to driver support, see
Paper 4.
Given that driving is considered as a cognitively motivated and controlled task,
it is possible to reorganise the four control layers in ECOM starting with the top
level of control targeting as an inner control loop and the monitoring, regulating
 5. Modelling the Driverin Control 99

FIGURE 5.8. An extended control

model adapted from Hollnagel's ex-
tendedcontrol model (Hollnagel and
Woods, 2005).

External events

and targeting loops as concentric circles with increasing diameters (see Fig. 5.8).
Thus, what Hollnagel described as the lowest level of control tracking will be the
outer circle representing the physical interaction with the interface to the vehicle 's
physical controls (e.g. steering wheel, pedals or various driver support systems).
Control goals are determined in the inner control loops and applied in the outer
control loops. That is to say that, for example in the targeting control loop the
goals are determined for the monitoring control loop. This flow is represented by
the outward-bound arrow-labelled goals at the top of Fig. 5.8. Feedback used to
modify and supervise the control is fed back from outer circles to inner control
circles represented by the in-bound arrow-labelled feedback at the top of Fig. 5.8.
The interaction between the driver and the physical environment is represented by
the two arrows-labelled external events and actions at the bottom of Fig. 5.8.
The driver can conduct an imaginary drive without physical interaction on the
targeting level. The driver can also, while driving , anticipate the possible outcome
of possible actions and let that determine the choice of action . However, even if the
cognitive skills are important, it is the lower level psychomotor skills that are the
foundation for safe driving (Ranney, 1997). Thus both cognitive and psychomotor
skills are needed for safe driving and furthermore the integrations and interaction
between the skills. The driver's mental model is, to a large extent, formed by the
experiences from driving - ' learning by doing' . Driver support can be provided
at different layers or levels of control and it seems likely that the ECOM can be
used to analyse and better under stand how different driver support systems can
influence driving behaviour.
100 Peters, VTI and Nilsson, VTI

5.12 Joystick Controlled Cars - An Example

Experiences from experiments with four-way joystick controlled cars for drivers
with severe disabilities will exemplify the application of the model discussed. The
joysticks were used to control speed (accelerating and braking) and steering and
thus replaced the standard controls (pedals and steering wheel). A manoeuvre test
on a closed track was performed with experienced drivers of joystick-controlled
cars (Ostlund and Peters, 1999). The test included the following three manoeuvres:
(1) firm and controlled braking on straight road, (2) firm and controlled braking
in a narrow curve, and (3) double-lane change. It was found that the test drivers
found it difficult to perform the straight road brake manoeuvre smoothly, which
was attributed to the lack of feedback. Braking in a curve was a manoeuvre, which
revealed problems of control interference between steering and braking control.
Finally, time lags in the steering control system caused problems, specifically when
performing the double lane-change manoeuvre.
The lack of tactile feedback when braking on a straight road was something that
mainly influenced the activities at the lowest level, tracking. The drivers were not
able to adjust the force applied on the brake lever in order to make a soft stop.
It can be speculated that the feedback they experienced, as whole-body g-forces
together with the visual cues, was not sufficient for the brake control function.
However, it is probably a delicate task to implement feedback forces to suit the
category of drivers who participated in the test, for example, drivers with sever
muscular dystrophy. Adaptation companies installing these systems often claim
that these drivers are extremely weak that active force feedback cannot be used.
However, this is probably not true because all manual control depends on some
form of tactile feedback. Whether it is technically feasible is another issue. In that
case it might be possible to investigate other sources of feedback, for example,
auditory feedback. However, such feedback will be artificial and not intuitive in
the same way as force feedback would probably be.
The interference problem observed during curve braking mainly affected the
activities on the regulating and monitoring levels. The drivers probably had a
mental model of how their joysticks worked in terms of steering and braking but it
was not sufficient to know what direction to move the lever in order to brake without
affecting steering control. They had to regulate and monitor their control actions
closely in order to adjust the joystick motion. This type of manoeuvre needs to be
carried out at least partly as anticipatory control. With increased experience the
drivers are likely to develop motor control schemata comparable to the precognitive
control proposed by McRuer et al. (1977). It was also observed that the drivers
compensated for performance decrements by driving slowly in the curve.
Finally, the time lag problem observed in the double-lane change had to be
handled as anticipatory control. The drivers initially had to plan (target) in advance
how to move the lever in order to compensate for the time lags and maintain control
in the double lane change. It seems very difficult to develop a mental model of
time lags and requires a lot of experience and training. It also turned out that the
double lane change manoeuvre was the most difficult task to carry out correctly
 5. Modelling the Driver in Control 101

(i.e., without hitting any cones). The joystick drivers performed worse than a group
of drivers matched with respect to age and driving experience. These drivers drove
a standard car. The joystick drivers hit more cones and produced higher lateral
acceleration forces despite driving their own individually adapted cars. All of the
identified problems with these joystick systems can be described and analysed in
terms of time-based safety margins.
Practical experiences have shown that it is a tedious task to learn to drive with this
type of control system (often implemented as a electro-hydraulic system). Thus, in
a follow-up simulator experiment an alternative joystick design was tested (Peters
and Ostlund, 2005). Time lags had been made similar to what is found in ordinary
car controls (steering wheel and pedals), steering and speed control had been made
more independent and active feedback was provided in the joystick lever. It was
found that the reduced time lag contributed substantially to make it easy to learn
to drive with this joystick. The separation of steering and speed control did not
as clearly improve performance but contributed somewhat to improved control.
The active feedback was not sufficiently tuned according to the individual drivers.
Thus, it was only drivers with unimpaired upper limb functions who could benefit
from the feedback.
In the example above it is obvious that the drivers were in need of support as
they could not drive a standard production car due to their physical impairment
(i.e., they were paralysed in their lower limbs and had impaired function in their
upper limbs). However, it can well serve as an example on how to understand
potential problems related to specifically ADAS and guide on how to resolve
them. What becomes obvious is that a driver in control depends on timely, sufficient
and intuitive feedback. Furthermore, integrated control functions with a high risk
of interference should be avoided. The human controller can often learn how to
compensate even badly designed support systems but a well-designed system will
both facilitate learning and ensure that the driver stays in control even in a critical
situation.

5.13 Summary and Conclusion

Driving task and driver behaviour modelling can be valuable to determine driver
support needs. Modelling can also be useful to determine whether the objectives
with a support system have been reached or to look for possible aversive conse-
quences. A driver support system can have different purposes and a range of dif-
ferent consequences for driver behaviour. However, the main focus in this Paper is
on driver support and safety. It has been argued that driving should be considered
a cognitive task of control in context. Furthermore, the hierarchical structure of
both the driving task and driver behaviour has to be considered. The principles of
control theory can be used to understand the driver as an active controller. Context
has to be included and system boundaries have to be determined. To understand
complex phenomena like behaviour adaptation, it was proposed to apply a closed
loop open system view. Thus, the cognitive control cycle that can be linked to
102 Peters, VTI and Nilsson, VTI

basic human abilities needed for the driving task was proposed as a useful model
of driver behaviour. Driver behaviour is also guided by a set of concurrent goals.
The homeostasis principle seems to be useful to understand the interrelationship
between different goals and driver behaviour. Human control seems to be very
sensitive to time. Time determines the margins for action. The cognitive control
cycle was described in terms of time. Finally the control cycle was expanded to
also consider hierarchical layers of control and the interaction with the system to
be controlled. The overall aim should be to ensure that the driver is in full control
even when driver support systems are installed. Well-designed support systems
are easier to learn and safer to use.

References
Allen, T.M., Lunenfeld, H. and Alexander, G.1. (1971). Driver information needs. Highway
Research Record, 366, 102-115.
Ashby, W.R. (1956). An introduction to cybernetics. Chapman and Hall Ltd., London.
Carsten, O. (2005). From driver models to modelling the driver: What do we really know
about the driver? In L. Macchi, C. Re and P.C. Cacciabue (Eds.). Modelling driver
behaviour in automotive environments. Office of Official Publiations of the European
Communities, Ispra, Italy.
Carver, C.S. and Schreier, M.F. (1982). Control theory: A useful conceptual framework for
personality - social, clinical and health psychology. Psychological Bulletin, 92, 111-135.
Damasio, A.R. (1999). The feeling of what happens: Body and emotion in the making of
consciousness. Harvest Book Harcourt Inc., San Diego, CA.
Engstrom, 1. and Hollnagel, E. (2005). Towards a conceptual framework for modelling
driver' interaction with in-vehicle systems. In L. Macchi, C. Re and P.C.Cacciabue (Eds.).
Modelling driver behaviour in automotive environments. Office of Official Publications
of the European Communities, Ispra, Italy.
Flach, 1.M. (1999). Beyond error: The language of coordination and stability. In P.A.
Hancock (Ed.). Handbook of perception and cognition: Human performance and er-
gonomics. Academic Press, New York.
Fuller, R.G.C. (1984). A conceptualisation of driving behaviour as threat avoidance. Er-
gonomics, 27, 1139-1155.
Fuller, R. (2005a). Control and affect: Motivational aspects of driver decision-making. In
L. Macchi, C. Re and P.C. Cacciabue (Eds.). Modelling driver behaviour in automotive
environments. Office of Official Publications of the European Communities, Ispra, Italy.
Fuller, R. (2005b). Towards a general theory of driver behaviour. Accident Analysis and
Prevention, 37,461-472.
Gibson, 1J. (1966). The senses considered as perceptual systems. Houghton Mifflin, Bostan.
Gibson, 1.1. and Crooks, L.E. (1938). A theoretical field-analysis of automobile-driving.
The American Journal of Psychology, 51,453-471.
Godthelp, H., Milgram, P. and Blaauw, GJ. (1984). The development of a time-related
measure to describe driving strategy. Human Factors, 26, 257-268.
Groeger, 1.A. (2000). Understanding driving -Applying cognitive psychology to a complex
everyday task. Psychology Press, Taylor and Francis, Hove, UK.
Hatakka, M., Kesikinen, E., Gregersen, N.P., Glad, A. and Hernetkoski, K. (2002). From con-
trol of vehicle to personal self-control; broadening the perspectives to driver education.
Transport Research Part F, Traffic Psychology and Behaviour, 201-215.
 5. Modelling the Driver in Control 103

Hollnage, E. and Woods, D.D. (2005). Joint cognitive systems: Foundations of cognitive
systems engineering. Taylor and Francis, Boca Raton, FL.
Hollnagel, E. (2002). Time and time again. Theoretical Issues in Ergonomics Science, 3,
143-158.
Jagacinski, R.I. (1977). A qualitative look at feedback control theory as a style of describing
behavior. Human Factors, 19(4),331-347.
Jagacinski, R.I. and Flach, 1.M. (2003). Control theory for humans-Quantitative ap-
proaches to modelling performance. Lawrence Erlbaum, Mahwah, NJ.
Janssen, W.H. (1979). Routeplanning en geleiding: Een litteratuurstudie. Institute for per-
ception TNO, Soesterberg.
Lee, 1.D. (2006). Driving safety. In R.S. Nickerson (Ed.). Reviews of human factors and
ergonomics. Human Factors and Ergonomics Society, St. Monica, CA.
Mckenna, F.P. (1982). The human factor in driving accidents. An overview of approaches
and problems. Ergonomics, 25, 867-877.
Mcruer, D.T., Allen, R.W., Weir, D.H. and Klein, R.H. (1977). New results in driver steering
control models. Human Factors, 19,381-397.
Michon, J .A. (1985). A critical view of driver behavior models: What do we know, what
should we do? In L.A. Evans and R.C. Schwing (Eds.). Human behavior and traffic safety
(pp. 487-525). New York: Plenum, New York.
Michon, 1.A. (Ed.) (1993). Generic intelligent driver support-A comprehensive report on
GIDS. Taylor and Francis, London.
Neisser, U. (1976). Cognition and reality: Principles and implications ofcognitive psychol-
ogy. W.H. Freeman and Company, New York.
Nilsson, L., Harms, L. and Peters, B. (2001). The effect of road transport telematics. In P-E.
Barjonet (Ed.). Traffic psychology today (1st ed., Vol. 1, pp. 265-285). Norwell: Kluwer
Academic Publishers.
Naatanen, R. and Summala, H. (1974). A model for the role of motivational factors in
driver' decision making. Accident Analysis and Prevention, 6, 243-261.
Ostlund, 1. and Peters, B. (1999). Joystick Equipped Cars' Maneuverability. VTI Medde-
lande No. 860A.
Peters, B. and Ostlund, J. (2005). Joystick controlled driving for drivers with disabil-
ities - A driving simulator experiment. Statens Vag och transportforskningsinstitut,
Linkoping,
Powers, W.T. (1998). Making sense ofbehavior - The meaning ofcontrol. Benchmark, New
Canaan, CT.
Ranney, T.A. (1994). Models of driving behaviour: A review of their evolution. Accident
Analysis and Prevention, 26, 733-750.
Ranney, T.A. (1997). Good driving skills: Implications for assessment and training. Work,
8,253-259.
Rasmussen, 1. (1986). Information processing and human-machine interaction: An ap-
proach to cognitive engineering. North-Holland, New York.
Reid, L.D. (1983). Survey of recent driving steering behaviour models suited to accident
investigations. Accident Analysis and Prevention, 15, 23-40.
Rumar, K. (1988). Collective risk but individual safety. Ergonomics, 31, 507-518.
Sanders, M.S. and Mccormick, EJ. (1993). Human factors and the automobile. In M.S.
Sanders and EJ. Mccormick (Eds.). Human Factors in engineering and design (7th ed.).
McGraw-Hill, New York.
Summala, H. (1985). Modeling driver behavior: A pessimistic prediction? In L. Evans and
R.C. Schwing (Eds.). Human behavior and traffic safety (pp. 43-65). New York: Plenum.
104 Peters, VTI and Nilsson, VTI

Summala, H. (1988). Risk control is not risk adjustment: The zero-risk theory of driver
behaviour and its implications. Ergonomics, 31, 491-506.
Taylor, D.H. (1964). Drivers' galvanic skin response and the risk of accident. Ergonomics,
7,439-451.
Vaa, T. (2005). Modelling driver behaviour on basis of emotions and feelings: Intelligent
transport systems and behavioural adaptations. In L. Macchi, C. Re and P.C. Cacciabue
(Eds.). Modelling driver behaviour in automotive environments. Office of Official Pub-
lications of the European Communities, Ispra, Italy.
Van Der Hulst, M. (1999). Adaptive control of safety margins in driving. Department of
Psychology. University of Groningen, Groningen, The Netherlands.
Weinberg, G.M. and Weinberg, D. (1979). On the design of stable systems. Wiley, New
York.
Weir, D.H. and Chao, K.C. (2005). Review of control theory models for directional and speed
control. In L. Macchi, C. Re and PC. Cacciabue (Eds.). Modelling driver behaviour in
automotive environments. Office of Official Publications of the European Communities,
Ispra, Italy.
Wickens, C.D. (1992). Engineering psychology and human performance. New York, Harper
Collins.
Wiener, N. (1954). The human use ofhuman beings. Da Capo Series in Science. Pergamon
Press, Oxford.
Wierwille, W.W. (1993). Visual and manual demands of in-car controls and displays. In
B. Peacock and W. Karwowski (Eds.). Automotive ergonomics (pp. 229-320). London:
Taylor and Francis.
Wilde, GJ.S. (1982). The theory of risk homeostasis: Implications for safety and health.
Risk Analysis, 2, 209-225.
6
From Driver Models to Modelling
the Driver: What Do We Really Need
to Know About the Driver?
OLIVER CARSTEN

6.1 Introduction
The variety of models of the driving task is almost as numerous as the num-
ber of authors who have contributed the models. Part of this variety is due
to the different applications for which the models are intended and another
part is due to the part of the driving task they are intended to describe. Since
driving encompasses so many tasks and subtasks at different levels, often per-
formed by the driver simultaneously, it is perhaps no surprise that it is hard
to find any consensus in the literature on how the process of driving should be
modelled.
Here the focus is on creating a structured model that can be used in real time, in
particular by a driver assistance system to monitor driver state and performance,
predict how momentary risk is changing, anticipate problem situations and in
response to adjust the behaviour of in-vehicle information systems and driver
assistance systems and also adjust feedback to the driver. The driver model would
therefore be the major component of a larger model supervising the interaction
among driver, vehicle and the traffic and road environment. The starting point
is a review of existing models to identify elements that can be used to predict
momentary risk.
The literature on models of the driving task is very extensive, going back at least
as far as Gibson and Crooks' Field of Safe Travel Model (Gibson and Crooks,
1938). Yet, in spite of the considerable effort put into producing such models, few
of them have been validated as predictive tools, apart from specific and limited
aspects of the driving task, where for example mathematical representations of
car following are used in microsimulation models of traffic. Still less is it the
case that any generalised model has been used in a driving assistance system to
guide the operation of the system as it supports the driver. Yet the potential of an
assistance system that 'understood' the driver is huge: it could give feedback to
novices, assist elderly drivers in difficult situations, inform a driver when he or she
is fatigued and adapt the operation of the vehicle to the needs of each individual
driver.

105
106 Carsten

6.2 A Typology of Models

Two broad types of driver model can be distinguished in the literature. The first
type is descriptive models. These models attempt to describe parts or the whole
of the driving task in terms of what the driver has to do. The second major type is
motivational models, which aim to describe how the driver manages risk or task
difficulty. The first type can be further subdivided into a number of categories -
there are task models, adaptive control models and production models.

6.3 Descriptive Models

These models attempt to describe either the whole of the driving task or some
element of it. A major feature of such models is that they are not predictive,
but are instead analytical. It is not possible to conclude from such models how
changes in driver motivation, capability or decision would affect the quality of the
performance of the driving task or situational risk.

6.3.1 Task Models

One major type of descriptive model of the driving process is the hierarchical task
model of Michon (1985). This model is shown in Fig. 6.1. The only conclusions
that can be drawn about performance or risk from this model is that a breakdown
of input, for example, environmental input at the control level or of feedback will
lead to a failure in the driving task.
Other more detailed task descriptions, such as the famous compilation of tasks
and sub-tasks by McKnight and Adams (1970), have similar qualities. They

Time Constant

General
Long
Plans

Controlled
Environmental secs
----+I Action Patterns
Input

Automatic
Environmental msec
----+I Action Patterns
Input

FIGURE 6.1. Michon's hierarchical structure of the road user task.

 6. From Driver Models to Modelling the Driver 107

d
External
Driver Disturbances
Remnant

Lateral
lateral Lateral Driver Steering - - - - - . ::~:I ....--.., Position, y
Driver Wheel Steering Steer
Position Position Operations
Operations Angle linkage Angle
Command on Functions
----+-.@~~ of Position
on Functions and Vehicle Heading
A ~y Error
of Heading Os Actuator 6 Angle.ljJ
YljJ
I Yy G
I
I
I
I
t

FIGURE 6.2. Compensatory model of driver steering (McRuer et al., 1977).

describe what the driver has to do in order to drive but they do not state ex-
plicitly how the quality of the driving performance is affected by the nature of
driver performing the tasks or by the nature of the driving being carried out. Risk
is only addressed implicitly in that a failure in performing a task is likely to lead
to a problem, but no guidance is afforded on what might cause such failure.

6.3.2 Adaptive Control Models

Another type of descriptive model describes the operation of the driving task in
terms of inputs, outputs and feedback. An example is the driver steering model of
McRuer et al. (1977) as shown in Fig. 6.2, where steering inputs are described as
compensation for errors in lateral position and heading angle. It can be noted that
this model does not account for non-deliberate errors in lateral position. Similar
models have been created for manoeuvring tasks. An example is shown in Fig. 6.3,
which is intended to describe the process of a driver approaching a T junction and
making a judgement about what speed to adopt in order to avoid collision with a
vehicle approaching from the right.
A more complex adaptive control model has recently been proposed by Holl-
nagel et al. (2003). This model covers both manoeuvring and control, as shown
in Fig. 6.4. Interestingly and presumably intentionally, this model covers only the
driver in control and not the driver who is losing or who has lost control.

6.3.3 Production Models

Tasks can also be described in terms of a formal set of rules, that is, as a production
system. Michon (1985) produced a formal set of such rules for changing gear in a
108 Carsten

Maintain
Speed

Brake
(Stop)

FIGURE 6.3. Driver decision making on approach to an intersection (Kidd and Laughery,
1964).

Anticipatory
control

FIGURE 6.4. Driver in control model (Hollnagel et aI., 2003).

 6. From DriverModels to Modelling the Driver 109

FIGURE 6.5. Action goal structure

@) for changinggear (Michon, 1985).
G=R G=1

@

G=2

@
G=3

(J)
G=4

four-speed car. The action goals structure for changing gear is shown in Fig. 6.5.
Here Action Goal 3, for example, shows the process of shifting into neutral on
stopping.
But once again such a model, while perhaps serving a purpose as a detailed
description of the purposes for which a driver changes gear, does not help to ex-
plain why one driver may choose a higher gear than another in identical traffic
circumstances. So it does not tell us in full why a driver changes gear at a partic-
ular moment. Car manufacturers know that motivation affects gear selection and
therefore have produced cars that feature automatic gearboxes that can change
style from sedate and economical to aggressive and sporty.

6.4 Motivational Models

Motivational models attempt to describe why drivers choose one alternative over
another in terms of utilities and trade-offs. In all probability, the most famous
motivation model of the driving task is the risk homeostasis model of Wilde (1982)
as shown in Fig 6.6. It is not the purpose here to provide a detailed critique of this
model - the literature is full of such critiques. More to the point is that Wilde
introduces the notion of driver capability affecting risk: perceptual skills, decision
skills and vehicle handling skills all feature in the model.
Other models introduce different aspects of driver motivation and capacity.
Naatanen and Summala (1974) add personality, experiences, motivation and vigi-
lance (see Fig. 6.7). Fuller (1984) adds capability and driver perception of rewards
and punishment (see Fig. 6.8). A more recent model by Fuller (2004) concentrates
110 Carsten

a
Expected Utilities of
Action Alternatives

4
Perceptual Skills

e
Resulting Accident
Rate
Lagged
Feedback

FIGURE 6.6. Risk homeostasis model of Wilde (1982).

on the interaction between task demands and driver capability, arguing that speed
choice in one major mechanism for the adjustment of task demands so that they
remain within capability. This latter model does not seek, however, to develop an
explanation of the major factors that determine capability; rather they are identified
in terms of broad groupings such as 'constitutional features' , that is, biological fac-
tors and 'human factors'. Focusing in particular on human factors, Rumar (1985)
describes the driving task in terms of information processing and introduces a
number of filters (physical, perceptual and cognitive) that can introduce errors. He
also incorporates such factors as motivation, experience, attention and expectation
in his model (see Fig. 6.9).

Personality Stimulus
Experiences Situation

Subj
Risk
Monitor

FIGURE 6.7. Risk threshold model

(Naatanen and Summala, 1974).
 6. From Driver Models to Modelling the Driver 111

FIGURE 6.8. Risk avoidance model (Fuller, 1984).

Reaction
Behavior

..., I

Physical Perceptual Cognitive FIGURE 6.9. Model of driv-

Filtering Filtering Filtering
ing as information processing
(Rumar, 1985).
112 Carsten

From an overall perspective, it can be seen that the motivational models introduce
more factors that permit prediction than the descriptive models. Thus, in theory,
they should be more subject to parameterisation and verification. However, such
empirical testing of the models has not generally taken place (see, e.g. Ranney
(1994), who argues that motivational models have normally not even been fully
specified let alone tested), and thus most of them remain as constructs rather than
as entities leading to the generation of rules and mathematical relationships.

6.5 Towards a Real-Time Model of the Driver

6.5.1 What Type ofModel Is Required?

In designing a model, a basic requirement is to establish how the model is to be
used and what should be its output - presuming, that is, the objective is to create
a predictive model. A second, but related issue, is that of granularity - at what
level of detail should the model operate. Here hierarchical models describing the
major elements in the driving task, such as Michon's (1985) can be useful. On the
whole, decisions related to the strategic level can be excluded as not relevant to
ADAS nor really important in determining momentary risk, although preventing
an impaired driver from setting out on a journey or even general trip suppression
or modal shift can reduce risk (see, e.g. Rumar, 1999). At the most detailed and
automated level of operation, namely vehicle control, we are faced with virtual
impossibility of predicting from one very small time step to the next what the
driver is likely to do. Such prediction would require almost absolute knowledge
about how the driver is controlling the vehicle, which would have to be based on
interpretation of physiological responses. Such prediction is beyond the capability
oftoday's cognitive science and may well be an impossible dream: humans are not
automatons whose precise reaction to a given complex situation can be predicted
with any certainty. A misinterpretation of a driver's decision could lead to severe
problems, for example preventing a lane change when the driver in the adjacent
lane has signalled consent to the manoeuvre.
This does not mean that control-level information would not provide a useful
input to the model. Thus, driver control of the vehicle could be analysed in real time
to provide the model with a depiction of a driver's characteristics. Such information
would be useful in 'tuning' adaptive systems to driver preferences. The AIDE
project has recently investigated adaptive forward collision warning systems, with
one type of adaptation being to observed driver reaction time, so that drivers who
habitually reacted quickly got later and hence less irritating warnings.
There is a strong argument, then, for focusing mainly (but not entirely) on the ma-
noeuvring level of driving. This is where many of the major risks occur - in decision
making at junctions, interactions with pedestrians, lane changing and overtaking,
etc. Errors in performing such manoeuvres are a major factor in accident causation,
so that in Great Britain 59% of all injury road traffic accidents occur at a junction
(Department for Transport, 2005). But this still begs the question of what should
be predicted. There are strong arguments for a probabilistic approach as opposed
 6. From Driver Models to Modelling the Driver 113

to a deterministic approach. Rather than predicting precisely and reliably what a

driver will do at any moment - an endeavour almost certainly doomed to failure
because of the variability of human response both between and within individuals
- a model should attempt to predict the probability of error or failure and thus
current and future risk. For example, there may be strong indications that a driver
is about to undertake an overtaking manoeuvre that, given the road layout and
traffic situation, may be highly risky. Feedback to the driver in advance of actually
starting the manoeuvre (so that the manoeuvre can be discouraged or prevented)
may well be more useful than feedback once the manoeuvre is already under way
(in the hope that it can be safely aborted). Control-level behaviour could be used
to trigger such warnings. A driver intending to overtake is likely to position the
vehicle close to or even slightly over the centreline.

6.5.2 Grouping the Factors

Thus it can be argued that an intelligent driver assistance system does not need to
fully comprehend all aspects of how a driver performs the driving task; it merely
needs to know about those driver factors that can affect the risk of vehicle operation.
This, of course, presumes that the major goal of such a system is to reduce risk,
but that seems a fairly reasonable constraint. The task of creating a model then
becomes one of identifying those driver factors that can be used to predict safety-
related performance or behaviour and of combining those factors into a structured
'model' that can be filled with rules and parameters, that is, a model that can be
empirically tested.
For this purpose, the various factors proposed in the motivational models dis-
cussed above can be identified as follows:

From Naatanen and Summala (1974)

o Personality

o Experiences

o Motivation

o Perception

o Vigilance

o Desired action

From Fuller (1984)

o Capability

o Arousal

From Rumar (1985)

o Motivation

o Experience

o Attention

o Expectation

o Perception and cognition

From Hollnagel et al. (2003)

o Goals/targets/plans
114 Carsten

Situation assessment
o

From Fuller (2004)

o Constitutional features (this model is somewhat unique in including biological

characteristics such as physical strength and reaction time)

o Training and education

o Experience

These factors can be grouped into categories as follows:

Attitudes/ personality
o Naatanen and Summala: Personality

o Rumar: Motivation

Experience
o Naatanen and Summala: Experiences

o Rumar: Experience

o Fuller (2004): Training and education

o Fuller (2004): Experience

Driver state (impairment level)

o Naatanen and Summala: Vigilance

o Fuller: Capability

Task demand
o Fuller (1984): Arousal

o Fuller (2004): Task demands

Situation awareness
o Naatanen and Summala: Perception
o Naatanen and Summala: Vigilance

o Rumar: Attention

o Rumar: Perception and cognition

o Rumar: Expectation

o Hollnagel: Situation assessment

Intention
o Hollnagel: Goals/targets/plans

o Naatanen and Summala: Desired action

We thus end up with five major categories of driver capability, performance and
behaviour that are related to risk. They are as follows:

1. Attitudes/personality
2. Experience
3. Driver state (impairment level)
4. Task demand (workload)
5. Situation awareness

Biological capability will be omitted on the grounds that permanent incapacity

relates to only a small proportion of drivers (although there may be empirical
arguments for including some aspects of age-related incapacity).
 6. From Driver Models to Modelling the Driver 115

As regards intention, this is simply a requirement by the model to predict the

driver's desired decision at a given moment. For example, we might infer from
lane position and closing towards the lead vehicle that a driver is intending to
perform an overtaking manoeuvre. The intention at one time step will then inform
the model at the next time step. The precise granularity of the model operation
will need to be defined, but it would seem sensible to start with fairly coarse time
steps to reduce the precision required.
For each of these categories, it is possible to verify from the accident literature
that they have a major role in terms of accident risk. For example, in terms of the
first category, Stradling and Meadows (2000) have cited the role of the propensity
to commit aggressive violations as a predictor of accident risk. For experience,
the models developed by Maycock et ai. (1991) relating age and experience to
accident risk can be used as evidence. For driver state, there is the literature re-
lating alcohol impairment to accident risk, including the most recent reanalysis
of the Grand Rapids data (Hurst et al., 1994). For workload, we can look at the
literature relating demands from road layout to driver performance. For the impact
of situation awareness, we can go to the studies of the impact of mobile phone use
on driving (e.g., Bums et al., 2002).

6.5.3 A Proposed Structure

A verifiable model requires a structure so that relationships can be proposed and
tested. A proposed structure, with the various categories grouped into long term
(years, months and days), medium term (hours and minutes) and short term (sec-
onds and milliseconds), is shown in Fig. 6.10.
Creating a predictive model still begs the question of what is to be predicted.
Any model created with today's knowledge must, of necessity, be probabilistic
rather than deterministic. It is not possible to create a reliable prediction of an
individual drivers decisions - after all even a drunk drive makes some correct

Task demand
(workload)

Long-term Medium-term Short-term

FIGURE 6.10. A causal structure for the categories of the driver model.
116 Carsten

FIGURE 6.11. Relationship between categories of driver factors and risk.

decisions. In addition, given sensor technology, a model running in real time cannot
be omniscient about the environmental situation and may indeed make errors in
interpreting driver actions and capabilities. So it is sensible in using some kind
of risk factor, that is, the risk of a serious error or the risk of crashing as the
dependent variable. The risk factor could be used in real time to identify when a
driver assistance system might need to warn the driver or to intervene in order to
keep prevent performance from deteriorating drastically with consequent impact
on risk. A relationship with performance and risk is suggested in Fig. 6.11.

6.5.4 Verifying the Model

Merely proposing a model does not justify it; still less does it demonstrate that the
model is useful. Thus, Grayson (1997) has stated, 'Although it would be unchari-
table to suggest that many psychological models [of driving] are little more than
the current buzz-words enclosed in boxes and joined by lines with arrowheads, it
is nevertheless difficult to avoid feeling that the term "model" has become deval-
ued in recent times'. The model proposed here is framework and not yet a usable,
validated predictive tool. Parameters for the factors need to be developed, and
the inter-relationship of the various parameters hypothesised and tested both as
regards their form (linear, exponential or power) and as regards the strength of the
relationship. Given the complexity of the proposed model, the number of relation-
ships within it (13) and the number of potential parameters, this may seem like an
impossible task. But a major consideration is that, for testing purposes, the model
can be decomposed so that not all parts are tested and verified at once. And prior
work can be found to throw light on the relationships. Thus, for example there
is already empirical evidence on the relationship between personality and driver
state in the form of studies that have examined the relationship between sensation
seeking and alcohol impairment (McMillen et aI., 1989; Yu and Williford, 1993).
Once the model has been verified or altered in parts, it should be recomposed
and tested in full. Such a test of the whole model would, of course, be a substan-
tial task requiring large numbers of drivers to be observed over a considerable
 6. From Driver Models to Modelling the Driver 117

amount of driving. But the payoff would be the delivery of a truly intelligent
co-driver.
A major step in model development and verification is identifying candidate
parameters to be included in each of the categories. Here one can conceive of a
two-stage process in which first of all the model is developed as an offline tool and
then at a second stage an online, real-time version of the model is created to run on
board the vehicle as an intelligent co-driver, supervisor of interaction with IVIS
and ADAS systems and manager of those systems to adapt to the current driver
state and current situation.
The second stage is clearly quite ambitious: care would have to be taken to
ensure that such a model did not create safety risks by going into unanticipated
states or unstable loops. One approach to minimise this would be setting boundaries
on system flexibility. Another, more straightforward alternative for an on-board
version is the creation of a simple rule-based structure that echoes the 'full' model
but does not emulate it.
However, the potential ambition of the second stage does not negate the use-
fulness of the first stage. It could have wide application and utility as an offline
design tool for new in-vehicle systems (creating 'what if?' scenarios, i.e. a kind
of failure mode and effects analysis), as a tool for evaluating road designs (e.g. in
safety audit) and even as a component in micro simulation models. In the last role,
it could replace the current rather crude practice of representing driver behaviours
by means of sampling from a set of built-in distributions, so as to represent for
example desired speed, typical headway or reaction time.
The model can initially be populated with relationships derived from the lit-
erature. Thus, since Maycock et al. (1991) have quantified the relationship be-
tween, on the one hand, experience and age and, on the other hand, risk, it may be
feasible to introduce some rules derived from that study into the model. However,
what is really required is the confirmation of parameters, conditional relation-
ships and interactions from empirical studies designed to test the hypothesised
relationships.

6.6 Developing an Online Model

A further implication of the categories is that if they are to used by a real-time
on-board system, then appropriate sensors will be required to generate information
about the current conditions, that is, the 'ifs' of the various rules. Some of that
information, for example, about driver age and experience could come from a smart
driving licence. But most of the required information will need to be observed by
the vehicle itself from how that vehicle is being driven. Thus, driver state and
appropriate methods for capturing it in real time has been investigated in a host
of European projects. However, Karel Brookhuis, one of the prime movers in this
research, has recently observed that' a valid framework for the evaluation of driver
impairment is still lacking' (Brookhuis and de Waard, 2003). Steering movements,
118 Carsten

with an increase in amplitude and a reduction in frequency indicating impairment,

are promising but to date a sufficiently reliable algorithm is lacking.
Estimating task demand in real time in order to manage driver workload has been
the focus of such projects as COMUNICAR (Amditis et al., 2002) and CEMVO-
CAS (Bellet et al., 2002). Map-based information on road type and layout can
provide basic essential information, for example on the frequency of intersections.
The fact that a driving is manoeuvring can be interpreted from vehicle yaw. Traffic
density can be provided from radar or image processing. Secondary task demand,
as opposed to driving task demand, can be inferred from interaction with enter-
tainment systems, navigation systems and other in-vehicle devices. Usage of the
mobile phone by the driver can be identified provided that there is an interface
between vehicle and mobile phone.
As regards situation awareness, the results of the HASTE project suggest that
it may be feasible to capture situation awareness from vehicle control parameters.
HASTE has been examining the impact of distraction from the use of in-vehicle
information systems on the driving task, and has identified very different impacts
from visual distraction as opposed to cognitive distraction (Carsten, 2004). Visual
distraction leads to increased lateral deviation, whereas cognitive distraction leads
to an apparent improvement in steering performance accompanied by an increase
in gaze concentration to the road straight ahead and loss of general situation aware-
ness. Thus a combination of eye movement cameras and steering sensors may be
able to provide sufficient indication of distraction and hence of loss of situation
awareness.
Steering behaviour can also potentially be used as an indicator of driving ex-
perience. Novice drivers have a generally reactive steering behaviour, whereas
experienced drivers have a more feed-forward strategy in which steering adapts to
road layout and obstacles (Jamson, 1999).
This leads to another problem that needs to be addressed. Steering behaviour
is clearly highly diagnostic, but we cannot use the same indicator, e.g. standard
deviation of lateral position, simultaneously to identify both visual distraction
and drink-driving. There is a need to investigate steering behaviour at a very
microscopic level to create unambiguous indicators. More refined indicators such
as steering reversal rate (the number of changes in steering wheel direction per
minute with a specific minimum threshold of movement) and steering entropy
(Boer, 2001) will be needed.
Finally, there is a need to identify driver intention. Salvucci and Liu (2002) have
shown how hard it is to identify a lane change that is about to be performed from
steering patterns and/or eye movements. They conclude that there is too much
variability in the data to be able to apply such methods reliably. However, there are
simpler methods for achieving this, and one that may well be effective as a signal
of a lane change is the use of the indicators by the driver. Furthermore, it could
be argued that, as in the overtaking case discussed earlier, intention need only be
known in a general rather than a precise sense. It may be more important and more
useful to know in sufficient time to deter the actual manoeuvre that a driver is highly
 6. From Driver Models to Modelling the Driver 119

likely to overtake than to detect the overtaking manoeuvre once it has started and
is too late to prevent. Here a system forearmed may mean a driver forewarned.

6.7 Conclusions
There are many driver models that are purely descriptive as opposed to being
predictive. Even the motivational models tend to be incomplete, addressing only
some of the driver factors that can elevate risk. A new structure has been proposed
here for a model that can be both verified and, if confirmed, applied in the long run to
monitor in real time the risk associated with the behaviour and performance of the
driver and to adjust feedback to the driver and tune the response of driver assistance
systems accordingly. It can be argued that, in order to produce a well-designed
advanced driver assistance system, particularly a complex multi-functional one,
such modelling is not only feasible but maybe even necessary. Otherwise, users
may reject such systems because they are not adapted to their needs.

References
Amditis, A., Polychronopoulos, A., Belotti, F. and Montanari, R. (2002). Strategy plan
definition for the management of the information flow through an HMI unit inside a car.
In Proceedings of the e-Safety Conference. Lyon.
Bellet, T., Bruyas, M.P., Tattegrain- Veste, H., Forzy, J.F., Simoes, A., Carvalhais, 1., Lock-
wood, P., Boudy, J., Baligand, B., Damiani, S. and Opitz, M. (2002). "Real-time" analysis
of the driving situation in order to manage on-board information. In Proceedings of the
e-Safety Conference. Lyon.
Brookhuis, K.A. and De Waard, D. (2003). On the assessment of criteria for driver im-
pairment: In search of the golden yardstick for driving performance. In Proceedings of
Driving Assessment 2003, the 2nd International Driving Symposium on Human Factors
in Driver Assessment, Training and Vehicle Design. Park City, Utah.
Boer, E. (2001). Behavioral entropy as a measure of driving performance. In Proceedings
of Driving Assessment 2001, International Driving Symposium on Human Factors in
Driver Assessment, Training and Vehicle Design. Aspen, Colorado.
Bums., P.C., Parkes, A., Burton, S., Smith, R.K. and Burch, D. (2002). How dangerous is
driving with a mobile phone? Benchmarking the impairment to alcohol. TRL Report 547.
TRL, Crowthome, UK.
Carsten, O. (2004). Implications of the first set of HASTE results on driver distraction. In
Behavioural Research in Road Safety 2004: Fourteenth Seminar (pp. 100-109). Depart-
ment for Transport, London.
Department for Transport. (2005). Road casualties Great Britain 2004. The Stationery
Office, London.
Fuller, R. (1984). A conceptualization of driving behaviour as threat avoidance. Ergonomics,
27, 1139-1155.
Fuller, R. (2004). Towards a general theory of driver behaviour. Accident Analysis and
Prevention, 37,461-472.
120 Carsten

Gibson, J.1. and Crooks, L.E. (1938). A theoretical field-analysis of automobile driving.
American Journal of Psychology, 51,453-471.
Grayson, G.B. (1997). Theories and models in traffic psychology - A contrary view. In
T. Rothengatter and E. Carbonell Vaya (Eds.). Traffic and transport psychology: Theory
and application. Pergamon, Amsterdam.
Hollnagel, E., Ndbo, A. and Lau, LV. (2003). A systemic model for driver-in-control. In
Proceedings of Driving Assessment 2003, the 2nd International Driving Symposium on
Human Factors in Driver Assessment, Training and Vehicle Design (pp. 87-91). Park
City, Utah.
Hurst, P.M., Harte, D. and Frith, W.1. (1994). The Grand Rapids dip revisited. Accident
Analysis and Prevention, 26, 647-654.
Jamson, A.H. (1999). Curve negotiation in the Leeds driving simulator: The role of driver
experience. In D. Harris (Ed.). Engineering psychology and cognitive engineering (Vol.
3, pp. 351-358). Ashgate, London.
Kidd, E.A. and Laughery, K.R. (1964). A computer model of driving behaviour: The
Highway Intersection Situation. Report VI-1843-V-I, Cornell Aeronautical Laboratories,
Buffalo, New York.
Maycock, G., Lockwood, C.R. and Lester, IF. (1991). The accident liability of car drivers.
Research Report 315. Transport and Road Research Laboratory, Crowthorne, UK.
Michon, lA. (1985). A critical review of driver behaviour models. In L. Evans and R.G.
Schwing (Eds.). Human behavior and traffic safety (pp. 485-520). Plenum Press, New
York.
McKnight, AJ. and Adams, B.B. (1970). Driver education task analysis. Vol. I: Task de-
scriptions. Human Resources Research Organization, Alexandria, Virginia. Final Report,
Contract No. FH 11-7336.
McMillen, D.L., Smith, S.M. and Wells-Parker, E. (1989). Brief report: The effects of
alcohol, expectancy, and sensation on risk-taking. Addictive Behaviors, 14, 477-483.
McRuer, D.T., Allen, R.W., Weir, D.R. and Klein, R.R. (1977). New results in driver steering
control models. Human Factors, 19,381-397.
Naatanen, R. and Summala, H. (1974). A model for the role of motivational factors in
drivers' decision-making. Accident Analysis and Prevention, 6, 243-261.
Ranney, T.A. (1994). Models of driving behaviour: A review of their evolution. Accident
Analysis and Prevention, 26, 733-750.
Rumar, K. (1985). The role of perceptual and cognitive filters in observed behavior. In
L. Evans and R.G. Schwing (Eds.). Human behavior and traffic safety (pp. 151-165).
Plenum Press, New York.
Rumar, K. (1999). Transport safety visions, strategies and targets: Beyond 2000. European
Transport Safety Council, Brussels.
Salvucci, D.D. and Liu, A. (2002). The time course of a lane change: Driver control and
eye-movement behaviour. Transportation Research Part F, 5, 123-132.
Stradling, S.G. and Meadows, M.L. (2000). Highway code and aggressive violations in UK
drivers. In Global Web Conference on Aggressive Driving. http://www.aggressive.drivers.
com/papers/stradling-meadows/stradling-meadows. pdf.
Wilde, G.1.S (1982). The theory of risk homeostasis: Implications for safety and health.
Risk Analysis, 2, 209-225.
Yu,1. and Williford, W.R. (1993). Alcohol and risk/sensation seeking: Specifying a causal
model on high-risk driving. Journal ofAddictive Diseases, 12(1),79-96.
 III
Learning and Behavioural Adaptation
7
Subject Testing for Evaluation of Driver
Information Systems and Driver
Assistance Systems - Learning Effects
and Methodological Solutions
KLAUS BENGLER

SUMMARY
Nowadays, subject testing represents a well-established methodology to evaluate
different properties of driver information systems and driver assistance systems.
Among several criteria, learnability is one important system property. User and
usage strategies are dependent on the subject's learning state, for example, to
switch attendance between driving task and operation of a driver information
system. Therefore it is wishful that the user acquires a model of the system, for
example learns as quickly as possible. Also, the intended usage of driver assistance
systems in given driving situations is influenced by the user's experience. A suitable
way to investigate related questions is to conduct a typical learning experiment
and to analyse data with the given methodology. In this type of experiment, the
familiarity and training state of the subject are set as independent variable. Beside
learnablity, other properties of human-machine interaction are to be investigated
and evaluated. In this case, however, the learning effectuated by a subject is an
important dependent variable or even noise in sense of measurement theory that
might cover a given main effect. After some empirical examples, possible solutions
will be discussed that help to manage this problem with justifiable expense.

7.1 Introduction
Following the idea of user-centered system design, subject testing is an estab-
lished and frequently used procedure to evaluate and ensure quality of human-
machine interface (HMI) concepts. Appropriate methods are applied in areas of
In-vehicle information systems (IVIS)l and advanced driver assistance systems
(ADAS 2 ) (Mayser, 2002). A good overview of different methodologies and of

1 Also called driver information systems (DIS).

2 Also called driver assistance systems (DAS).

123
124 BENGLER

their interaction with possible product development processes is given in the pro-
ceedings of Bundesanstalt fur StraBenwesen (2000).
Documents like DIN 66234 Part 8 (DIN66234 1986) and the European Statement
of Principles (2000/53/EC 2000) make clear that beneath properties like 'error
robustness', 'interruptability' or 'visual demand', learnability of a system is also
an important system property. More and more questions are raised concerning the
quality of subject testing in the sense of test theory and methodology (Kanis, 2000;
Haigney 2001). This paper discusses the conditions under which an independent
evaluation of dialogue quality and assistance characteristics can be conducted. An
evaluation that accounts for the fact that learnabilty and learning state should not
influence the results erroneously.
Frank and Reichelt (2001) also mention that learnability is one criterion that
can be tested during the development of an ADAS by expert judgement and ex-
perimental testing to ensure high product quality and system acceptance.
In general, learning is defined as a 'permanent change of behaviour based on
experience' (Hilgard and Bower, 1966) and therefore parameters for the learning
process were modelled for given tasks. More specifically, Woods (1999) states that
the user is building up heuristics and simplified models to structure the interaction
with a given system.
In addition, Reeves (1999) introduces the concept of 'cognitive complexity'
as quality for an HMI based on models for cognitive processes. Cognitive com-
plexity describes the property of a system that enhances or prohibits the user's
learning process. This process includes elements of perception, model building
and categorisation. Following this, Reeves (1999) gives recommendations for sys-
tem interaction concepts and information presentation to increase the usability
of an HMI concept. These include guidelines for a leamer-oriented design pro-
cess that targets interaction concepts that support perception, visualisation, model
building, categorisation and problem solving (cf. also Groeger, 1991).
This emphasises that learnability of a system is a predominant feature of an
interaction concept besides other properties that contribute to usability and utility
of the system. Therefore, in the following, learning experiments as one type of
experiment will be distinguished from system evaluations.
A learning experiment typically focuses on questions about acquisition of
knowledge on the system and users' mental models. Questionnaires and inter-
views are used as dependent variables as well as performance measures (e.g.
driving performance, user errors). In case of a system evaluation, properties like
visual or motoric demand are analysed as well as additional performance measures
and user errors - that is observable behaviour (Woods 1999) describes the idea
of so-called built in or designed system diseases as one of the most important
source for human errors in contrast to the mostly stated human error and give more
information on error analysis as a further evaluation method. Thus, in both types
of experiments, the analysis is based on similar measurements. If a subject was
learning how to operate the system during a system evaluation, the learning pro-
cess will produce variability and cover the main effect and item of interest of this
investigation.
 7. Subject Testing for Evaluation of Driver Information Systems 125

In this context, subjects' learning capabilities and inter-individual differences

playa very important part of the learnability of the system. One must not oversee
that in most cases experiments and data gathering take place in an automotive
environment (vehicle or driving simulator) that is not familiar for the subject.
Some examples shall help to get a better understanding of the quality and range
of learning effects after some methodological discussions.

7.2 Methodological Issues

In general the in-car usage situation that is the learning context for a given in-car
HMI is described with the following model: The driver has to fulfil the primary
driving task that can be subdivided in stabilisation, manoeuvring and navigation.
While stabilisation includes highly automated activities on the skill level, manoeu-
vring and navigation afford more rule- and knowledge-based user behaviour. Also
these three subtasks and the acquisition of skills, rules and knowledge are sub-
ject to learning experiments. But the focus of the given publication is secondary
and tertiary tasks include the operation of driver information and driver assistance
systems like audio, navigation, HVAC and ACC.
Obviously primary, secondary and tertiary tasks use the same limited motoric,
sensory and cognitive resources of the driver. The design of in-car HMI therefore
requires that the primary task does not crucially interfere with a secondary or
tertiary task. Apparently all these tasks underly learning processes that contribute
to increasing driver performance. In the case of the primary task this is pointed
out by a decrease in accident risk with an increase of driving experience. The
content learnt by the driver is very complex as it includes procedural knowledge,
semantic knowledge as well as knowledge on system behaviour and system limits.
The systems that construct these learning 'items' range from the traffic system in
general over traffic situations, the vehicle in its static and dynamic properties up
to devices and functionalities.
To acquire this knowledge, drivers apply strategies that are well known from
other domains: especially, rules on the traffic system and basic interaction with
the vehicle. Learning strategies are learnt by education and observation and after
this refined by training/exercise. Operation of in-car functionalities is, to a high
degree, based on the application of transfer of trial-and-error strategies. Manuals
might also playa minor role.
Manstetten (2005) emphasises the concept of learnability of driver assistance
systems:

With reference to the project RESPONSE, 'a system is learnable, if accurate

assimilation of information by the driver occurs, evidenced in the driver's un-
derstanding of system function, system handling and situational limits ' .
A 'self-explanatory' support system is defined as a 'driver assistance system
leaving a minimal amount of learning demand to the driver and eliminating
learnability issues which can result in safety-critical traffic situations' .
126 BENGLER

This makes clear that learning processes expand the range of the driver especially
by the establishment of highly automatised expert behaviour. Vehicle stabilisation
(distance and lane keeping) and the more efficient usage of in-car devices based on
usage strategies are an example. An example for such a strategy is to use a given
device not in any but a suitable traffic situation or to interrupt the operation and
continue afterwards (Sayer et al., 2005).
Learnability is therefore a beneficial system property that is to be tested
('Learnability' DIN 66234/8) like other ergonomic qualities during product de-
velopment. A problematic fact in this context is that only few models ex-
ist' that describe the learning behaviour described in the section above. This
fact makes it difficult to plan and conduct learning experiments during system
development.
On the other hand, learning is also a potential source of 'noise' in empiri-
cal testing to other values describing HMI qualities like visual demand or in-
terruptability. Therefore, learning processes have to be taken into account ei-
ther as a main effect or as a side effect at the different steps of the evaluation
process:

planning (sample construction, subject selection, procedure, scenarios);

conduction (training procedure, training criterion, measurements);
analysis (data qualification, interpretation, explanation).

This is now the point to stress the lack of detailed learning models that could
help the experimenter to decide on the above questions.
Especially the model of Rasmussen (1983) is suited and used to describe learning
in the context of the driving task.
In relation to the driving task, learning is described as a staged process
Rasmussen (1983), however, the transition between skill-based - rule-based -
knowledge-based stages are not described very concisely. Therefore, the stages
are difficult to handle for planning and experimental procedures. But they can be
used post hoc to reduce variability, in most cases, as a covariate.
The learning processes in relation to in-car HMI are mostly modelled using
the power law. This shall also help to deal with these effects within experimental
procedures. As different system types and usage scenarios might require a very
differentiated discussion, mostly the necessary parameters to describe the power
law function are missing.

7.3 Experimental Examples

The following section will give examples for learning experiments and show that
there is a possibility to describe in-car learning behaviour. Established methods can
be used for this purpose. The description and modelling of these specific learning
processes could then help to plan and conduct future evaluations of comparable
functionalities.
 7. Subject Testing for Evaluation of Driver Information Systems 127

7.3.1 Evaluation ofa Multimodal HMI

Strategies to switch between secondary task and driving task are highly dependent
on the learning state. Therefore, learning processes are an important factor during
the evaluation of driver information systems. This can be seen looking at the
evaluation of multimodal HMI demonstrator and error analyses on interaction
data (cf. Bengler, 2002).
Within a field test, learning of a multimodal HMI that allowed to operate a
navigation system and an in-car phone was investigated. Subjects could change
anytime during operation between input modalities (speech vs manual turn/push
knob). The study investigates user behaviour as well as driving behaviour in dif-
ferent traffic situations. Specific interest is put on learning process and effects that
are correlated with learning.
The questions were as follows:
Is the user able to operate typical driver information systems multimodally while
driving?
Is it possible to acquire necessary knowledge in appropriate time (i.e. to learn
an efficient way of multimodal interaction)?
The study was conducted as a 2 x 2 within-subject design with independent
variables secondary task (telephone/navigation) and primary driving task (sim-
ple/complex). The telephone task included to dial an instructed phone number
while driving. The navigation task required to enter a destination (city and street)
while driving. Subjects had free choice of the input modality (manual, voice). The
simple driving condition was driving on a straight one or two/lane road, with right
of way and only few traffic lights (the complex section was on narrow curvy urban
roads without lane markers including yield/pass situations).
To negotiate the complete test track (12 km) required about 25 min. All subjects
(N = 11) had a valid driving license, continuous driving experience and no or only
minimal experience with speech recognition systems.
After a very short introduction and a very short training using the speech recog-
nition system, each subject had to go through the test course three times: once
without and twice with secondary task operations, following the instructions of
the experimenter.
Data during voice dialling at the beginning and at the end of the experiment
were used to analyse learning effects.
Driving performance and in-car operations of the subjects were videotaped.
Considering the first question, all subjects safely negotiated the secondary tasks,
and compared to other studies, speed was not significantly reduced during sec-
ondary task operation.
The videotaped data were analysed following an error taxonomy that differen-
tiates the following:
Timing errors: (user speaks to early):
User speaks to non-active ASR
User speaks while system's speech output
128 BENGLER

40

35

30

25

20

15

10

o-J,.-"
Timing Vocabulary Dialog Recognition
Errors Errors Errors Errors

FiGURE 7.1. Frequency of different error types during first and last phases of experiment
(all subjects and traffic situations).

Vocabulary errors : Non-valid/available command word used; Irrelevant item

used (e.g. thinking aloud, laughing, hesitations 'aahm'); Inadequate command
(e.g. ' DELETE' deletes all digits, instead 'ERROR' deletes only last digit) .
Dialog errors : User ignores system prompt; User interprets system prompt in
the wrong way.
Recognition errors: ASR recognises a correct user input in the wrong way; ASR
'ignores' a correct user input.
The error analysis shows that there is a well-known but remarkably fast learning
effect that can be seen in the subjects ' error behaviour (cf. Fig. 7.1) .
A significant error reduction in all error categories can be seen . This effect is
not based on a changed mode of operation of the speech system under test, but
rather on a changed usage and speech behavior of subjects . This means that appro-
priate command words are spoken at the right point of time . Within the dialogue,
questions from the system are answered correctly by subjects . The reduction of
recognition errors refers to the fact that subjects corrected their level of speech
volume . It can be shown that on the one hand the number of successfully finished
dialogues increases due to a change of behavior, and on the other hand, a very
strong learning effect using speech-operated systems can be observed. In case of a
system evaluation using a treatment; - treatments design - comparing a speech-
operated HMI A to a second speech-operated HMI B - the second treatment under
test would profit the subjects' experience with the first treatment. Assuming that
both speech-operated systems would 'behave' the same way regarding vocabulary,
timing and dialogue.
Analysis of total task times presents a similar result. During the experiment, total
task times are remarkably reduced (Table 7.1). A more detailed analysis shows that
a reliable estimation of total task time is possible after the second run.
 7. Subject Testingfor Evaluation of DriverInformation Systems 129

TABLE 7.1. Totaltask times (mean and standarddeviation) telephone and navigationtasks
in four subsequentruns during the experiment (N == 11).
Telephone tasks Navigation tasks

2 3 4 2 3 4

Mean 104.4 40.7 38.9 41.0 56.7 35.6 40.2 36.0

SD 98.06 21.48 18.31 21.00 34.13 7.45 16.23 14.46

7.3.2 Destination Entry While Driving

A second experiment shall outline that learning effects have to be taken into ac-
count if dialogue properties like workload or demand that are often based on total
task time, total glance time or single glance time are evaluated. For these vari-
ables, precise criteria values are formulated for evaluation. The critical question
is whether only the 'worst case' of an novice user should be tested or also the
efficiency potential of a given dialogue performed by a trained user. Because the
goal should be that a dialogue is performed in the beginning, but performance
increase is given with increasing user experience and practice.
Destination entry shall serve as a secondary in-car task and an example for
a dialogue of long total task time. Within a learning experiment, Jahn (2002)
compared total task times of destination entry dialogues, using different navigation
systems. Of interest is whether a learning process takes place and which parameters
describe the learning process of this specific dialogue in the best way.
Driving on a test track, subjects were instructed to enter a destination consisting
of city and street. The system was operated using a turn push button.

Twelve subjects participated in the experiment (aging 35-55 years) having more
than 100,000-km driving experience.
In sum, 100 destination entries were negotiated at speeds of 40-50 km/h.
Two route complexities are distinguished.
Easy: 1.5 km, straight
Winding: 1.2 km; turning into narrow roads
The resulting data set gives insight into the learning process of this very specific
task.
At first, a learning process can be seen with a clear performance improvement
between Blocks 1 and 2 for the winding route. In addition, there is a significant
influence of route complexity on mean duration for destination entry and the level
of the learning process in general (see Fig. 7.2).
Comparison of different subjects shows two further effects. Learning processes
of subjects 1 to 5 start at significant different levels and are hardly comparable in
their shape. This means that either the mean value computed of this sample would
go along with a high standard deviation. This is especially problematic if the data
would be used for comparison with another sample using a different system. Or,
the experimenter would have to decide to extract subgroups of subjects due to their
difference in learning behaviour.
130 BENGLER

130 r---T-----r----r----r--~-___,____, FIGURE 7.2. Mean duration of a destination

120 entry depending on the experimental session
~ 110 and route characteristic. Block 1 = 20 desti-
~
c: 100 nation entries. Block 2-6 = at times 16 desti-
Q) 90 nation entries.
80
70
... -....-_.---- --.
60
50
40
30
20 winding - -
easy ---------
10
o '-----"_----'-_--'--_-'----------.L._---'-----'

2 345 6
Block

Figures 7.2 and 7.3 show that both effects -learning process and inter-individual
differences - result in a high level of variance of the data set. Therefore, they should
be treated as error variance in system evaluation experiment.

7.3.3 Evaluation ofDriver Assistance Systems

In the area of empirical evaluation of driver assistance systems, learnability and
learning effects are an important parameter. Simon and Kopf (2001) emphasise this
vividly in the context of automatic cruise control (ACC) systems. The study con-
ducted was a long-term field trial to investigate learning effects. Five subjects who
participated used an experimental car equipped with ACC for 3 weeks for their ev-
eryday trips. Data show that system experience increases with usage time. But also
that subjects dealing with the limits of the ACC system establish different strategies.

130
120
~ 110
e-
C 100
<D
90
en<D 80
"0
70
'0
c 60
0
~ 50
:; P1
"0
40 P2 ---------.
C 30 P3 -----------
ctj
<D 20 P4 FIGURE 7.3. Mean duration of a destination
~ P5 _._._._._--
10 entry depending on experimental run per
0 subject (PI to P5). Block I = 20 destination
2 3 4 5 6 entries. Block 2-6 = at times 16 destination
Block entries.
 7. Subject Testing for Evaluation of Driver Information Systems 131

The authors distinguish 'eager testers' and 'careful approachers'. Obviously, dif-
ferent usage strategies again lead to a remarkable variance within this small sample.
This variance can be used to describe different learning strategies. But driving per-
formance and other effects - especially at the beginning of the experiment - have
to be interpreted by taking account of these individual learning strategies.
An attempt to present a model on skill acquisition of an ACC system is given by
Hoedemaeker (1999). She proposes the idea that the adaptation of the ACC mental
model results of two processes that operate at different abstraction levels within
the mental model. The first process is based on the difference of expected system
behaviour provided by the user's mental model of the ACC and the observed system
behaviour. Depending on the degree of wrong predictions, the ACC mental model
is updated to arrive at a more accurate account of the ACC's operational domain.
The second process dealing with the adaptation of the mental model takes place
at a higher abstraction level. The intensity of ACC usage then is highly correlated
with the degree of perceived inconsistencies between ACC behaviour and user
expectation based on his or her mental model. The less inconsistencies the more
system usage and vice versa.
A very sophisticated long-term field operational test using an ADAS was con-
ducted by Weinberger (2001). The experiment gathered more information about
the learning process for the usage of controls and display and the judgement of
take-over situations. Participants used an ACC-equipped car for a period of 4
weeks per person. Data were analysed with respect to the duration of the learning
phase. The change in behaviour was investigated using drivers' self-assessment of
the length of learning and observed driving behaviour during take-over situations.
The results suggest that 2 or 3 weeks are needed to learn the operation of ACC and
the assessment of take-over situations for a goal-directed system usage. Interesting
is the methodological advice given by the authors that other ACC users might need
a different learning time as the participants in the study drove much more than the
average driver.
Kostka et al. (2004) recommend the analysis of driver errors by expert observa-
tion as a practical empirical method to evaluate workload and distraction but might
serve as well as a tool to investigate the learnability of the concept and especially
erroneous user expectations on system functionality.
Results from this study and that of Weisse (2002) give the impression that the
investigation of learning processes and individual behaviour in the context of driver
assistance systems demands considerably more effort than driver information sys-
tems.

7.4 Solutions
The discussion of the 'learning problem' and the selected experiments show that on
the one hand learning processes and learnability in the in-car domain can be inves-
tigated using classical methodologies. Due to this method, the effort is remarkable.
132 BENGLER

On the other hand it is also true that learning processes can lead to considerable
data variablity in evaluation experiments:

learning and carry -over effects;

positive or negative transfer in within-subjects designs treatment A, which pre-
cedes treatment B in the experiment, will have influence on the variability of
values in the second measurement; and
everyday knowledge and experience will interact with subjects' performance.

Unfortunately, it is not possible to compute or estimate error variance as expe-

rience and concise models for these effects are still missing. This procedure might
become feasible if the characteristic learning curve was known for a given system.
After the citations of learning experiments and related problems, some approaches
shall be presented to control learning effects in system evaluations.
Inter-individual differences should be covered most efficiently using a within-
subjects design. A systematic distribution of subjects on experimental treatments
using a Latin Square can avoid or at least balance positive and negative transfer
effects over the whole experimental data set. Using this approach, the resulting
number of subjects must not be underestimated (Bortz, 1999).
The most reliable solution of the learning problem is to use highly trained
subjects combined with the application of a standardised performance criterion
before measurement begins. This approach, however, requires a substantial effort
for training and testing. The study of Jahn et al. (2002) gives an impression of the
necessary training effort for a destination entry example.
A further possibility is to use so-called 'expert users' as subjects, for example,
for conduction of a destination entry study, a sample of experienced users who
have everyday experience with the given system, however, in this case also, a
performance test should be conducted. But especially for the investigation of novel
usage concepts this approach is no alternative.
A general recommendation on the conduction of learning experiments with
driver information systems is given by Rauch et al. (2004). They recommend that
learning experiments that investigate in-car systems for use while driving should
be conducted in a dual-task setting to increase the validity of the method.

7.5 Conclusions
The examples give an impression of learning effects that have to be expected
in connection with subject tests evaluating driver information systems and driver
assistance systems. It turns out that learning experiments that investigate technolo-
gies such as voice recognition and future driver assistance systems would ease the
planning and conduction of system evaluation and raise their quality.
In addition, one can state that the investigation of total task time and other
performance measures is only reliable and makes sense if learning is finished and
subjects are in a stable state.
 7. Subject Testing for Evaluation of Driver Information Systems 133

Suitable models are still missing that would describe learning of driver informa-
tion systems and driver assistance systems in detail. Therefore, the experimenter
has to reduce high variances that are based on learning effects by carefully selecting
experimental procedures that do not exceed manageable effort.

References
2000/53/EC (2000). 2000/53/EC: Commission recommendation of 21 December 1999 on
safe and efficient in-vehicle information and communication systems: A European state-
ment of principles on human machine interface. Official Journal of the European Com-
munities L19 43 (25.01.2000), 64-68.
Bengler, K., Noszko, Th. and Neuss, R. (2002). Usability of multimodal human-machine-
interaction while driving. In Proceedings of the 9-th World Congress on ITS, CD-Rom,
9th World Congress on Intelligent Transport Systems, ITS America, October 2002.
Bortz, J. (1999). Statistik fur Sozialwissenschaftler (5th ed.). Springer-Verlag, Berlin-
Heidelberg.
Bundesanstalt fur StraBenwesen (2000). Informations- und Assistenzsysteme im Auto be-
nutzergerecht gestalten. Methoden fur den EntwicklungsprozeB. Bd. Heft 152. Bergisch
Gladbach: Bundesanstalt fur StraBenwesen, 2000.
Norm DIN66234 8 8 (1986). Bildschirmarbeitsplatze: Grundsatze der Dialoggestaltung.
Frank, P. and Reichelt, W. (2001). Fahrerassistenzsysteme im Entwicklungsprozess. In Th.
Jurgensohn and K.-P. Timpe (Hrsg.) (2001). Kraftfahrzeugfuhrung. Springer, Berlin,
Heidelberg, New York, pp. 71-80.
Groeger, J.A. (1991). Learning from learning: Principles for supporting drivers. In Inter-
national Symposium on Automotive Technology & Automation (24th: 1991: Florence,
Italy). Croydon, UK: Automotive Automation, pp. 703-709.
Haigney, D. and Westerman, SJ. (2001). Mobile (cellular) phone use and driving: A critical
review of research methodology. Ergonomics, 44(2),132-143.
Hilgard, E.R. and Bower, G.R. (1966). Theories of learning (3rd ed.). Appleton-Century-
Crofts, New York.
Hoedemaeker, M. (1999). Driving Behaviour with Adaptative Cruise Control and the Ac-
ceptance by Individual Drivers. Delft University Press, Delf.
Jahn, G., Krems, J.F. and Gelau, C. (2002). Skill-development when interacting with in-
vehicle information systems: A training study on the learnability of different MMI con-
cepts. In D. de Waard, K.A. Brookhuis, J. Moraal and A. Toffetti (Eds.). Human fac-
tors in transportation, communication, health, and the workplace. Shaker Publishing,
Maastricht, The Netherlands.
Kanis, H. (2000). Questioning validity in the area of ergonomics/human factors. Er-
gonomics, 43(12), 1947-1965.
Kostka, M., Dahmen-Zimmer, K., Scheufler, I., Piechulla, W. and Zimmer, A. (2004).
Fahrfehlerbeobachtung durch Experten. Eine Methode zur Evaluierung von Belastung
und Ablenkung des Fahrers. In A.C. Zimmer, K. Lange, et al. (Hrsg.). Experimentelle
Psychologie im Spannungsfeld von Grundlagenforschung und Anwendung. Proceedings
43. Tagung experimentell arbeitender Psychologen (CD-ROM). Regensburg.
Manstetten (2005). Evaluating the traffic safety effects of driver assistance systems. AAET
2005, Automation, Assistance and Embedded Real Time Platforms for Transportation:
Airplanes, Vehicles, Trains, 6th Braunschweig Conference. (Vol. 1). Braunschweig, DE.
134 BENGLER

Mayser, C. (2002). An advanced concept for integrated driver assistence systems

(S.A.N.T.O.S.-Projekt). Proceedings of the 9-th World Congress on ITS, CD-Rom, 9th
World Congress on Intelligent Transport Systems. ITS America, October 2002.
Rasmussen, J. (1983). Skills, rules and knowledge; signals, signs and symbols and other
districtions in human performance models. IEEE Transactions on Systems, Man and
Cybernetics, SMC 13(3).
Rauch, N., Totzke, I. and Kruger, H.-P. (2004). Kompetenzerwerb fur Fahrerinformation-
ssysteme: Bedeutung von Bedienkontext und Menustruktur, In VDI-Berichte Nr. 1864.
Integrierte Sciherheit und Fahrerassistenzsysteme. VDI-Verlag, DUsseldorf.
Reeves, W. (1999). Learner centered design. Sage Publications,Thousand Oaks, CA.
Sayer, J.R., Devonshire, J.M. and Flannagan, C.A. (2005). The effects of secondary tasks
on naturalistic driving performance. Report No. UMTRI-2005-29. The University of
Michigan TransportationResearch Institute, Ann Arbor, Michigan, 48109-2150.
Simon, J. and Kopf, M. (2001). A concept for a learn-adaptive advanced driver assis-
tance system. In CSAPC 2001, 8th Conference on Cognitive Science Approaches to
Process Control, The Cognitive Work Process: Automation and Interaction, Universitat
der Bundeswehr, Neubiberg, Germany, 2001, pp. I-II.
Weinberger, M. (2001). Der EinflufJ von Adaptive Cruise Control Systemen auf das
Fahrerverhalten. Shaker Verlag, Aachen.
Weisse, J., Landau, K., Mayser, C and Konig, W. (2002). A user-adaptive assistance sys-
tem. In Proceedings, ORP2002 - 2nd International Conference on Occupational Risk
Prevention, February 2002.
Woods, D.D. and Cook, R.I. (1999). Perspectives on human error. Hindsight biases and
local rationality. In F.T. Durso (Ed.). Handbook of applied cognition. John Wiley and
Sons, Riverton, NJ.
8
Modelling Driver's Risk
Taking Behaviour
WIEL JANSSEN

8.1 Introduction
A realistic estimate of the risk-reducing effects that will be gained by the intro-
duction of advanced driver supports (ADAS) requires knowledge of a number of
elements, of which user behaviour is the least understood. This paper focuses on
some of the most essential knowledge that is already available, in particular on the
mechanisms by which users could possibly change their behaviour once they start
using the support.

8.2 Expected Risk Reductions from New Technology

on the Road
The safety effects resulting from the introduction of new technology in vehicles
or - for that matter - in the road infrastructure follow from the interaction of four
components:

1. The so-called 'engineering estimate' or intrinsic effectiveness estimate of a

device's expected safety effect, that is, the accident reduction to be expected on
the basis of purely statistical or mechanical considerations. For example, the
seat belt's effectivity in increasing the probability of surviving a vehicle crash
is estimated to be around 40% to 50%. This would then be the initial estimate
of the reduction in fatalities if the entire population would use the belt.
2. The degree of penetration or use rate of the device in the relevant population. For
devices which rely, for their effectiveness, on the acceptance by the population,
there is the issue of selective recruitment, which means that the use rate per se
and/or the effect a measure achieves is affected by self-selective processes in
the population. The hypothesis is that those who opt for the device differ from
those who do not in respects that are essential to measure the effectiveness, the
particular assumption being that those who are least inclined to accept a safety
device would profit the most from it (e.g. Evans, 1985).

135
136 Janssen

3. Effects on and changes in the user's behaviour that may be brought about by
the device, in particular the so-called behavioural adaptation processes.
4. The functional relationships linking behavioural parameters (e.g. driving speed
and its variability) to resultant accident probability and severity.

We need to know more on all of these, but this paper focuses on the last two
and on behavioural adaptation in particular. From the behavioural point of view,
selective recruitment is an almost equally important issue, but it deserves more
space than could be devoted here.

8.3 Behaviour When Driving with Supports

A number of behavioural factors are critical to the success of a support system

(Janssen, 2001).

8.3.1 The Importance of Plain Old Ergonomics

The first factor is the badly designed in-vehicle (or roadside) supports, which
may cause mental over- or under load of drivers, both of which may lead to
a decrease in situational awareness and to increased risks. Even recent, on-the-
market, systems sometimes suffer from this (Janssen et al., 1999). This should
no longer be acceptable, given the extensive knowledge on display and han-
dling characteristics, colour use, lettering, etc., that is contained in ergonomic
handbooks.

8.3.2 The Loss of Potentially Useful Skills

When drivers are relieved from executing certain elements of the driving task,
they may lose the ability to perform the associated skills. This is the equivalent of
certain everyday phenomena, like younger people in Western societies supposedly
no longer being able to do even simple additions without the help of a calculator.
That skills are lost may not matter at all in some cases and it is certainly not
necessary to retain each and every antique skill for some doomsday to come;
however, sometimes the skill that is gone could well have been a safety asset in
some situations.

8.3.3 Opportunities for New Errors

While advanced supports may take the opportunity away for drivers to make er-
rors, the design and the required maintenance of these systems may themselves
constitute new sources of errors. This also needs to be taken into account when
estimating net safety effects in, at least, a qualitative sense.
 8. Modelling Driver's Risk Taking Behaviour 137

8.3.4 Problematic Transitions

It is generally recognised that the introduction of (partially) automated supports
will require a solution for problems associated with the ensuing mix of vehicles
that have the support and those that have not. This is the problem of a transition
taking place in time.
Less attention has been directed to the problems a driver may experience when he
or she has to make a transition in space, that is, from non-automated to automated
parts of a road network and vice versa. In particular, adaptation and take-over
effects from the automated environment may play a negative role when getting
adjusted to the non-supported environment again.
On a more general level, the driver may actually have to learn dealing with
advanced supports. This may require fundamental changes in training curricula,
as they presently exist.

8.3.5 Risk and Risk Perception: My Risk and Yours

People accept less risk that is imposed on them by others than when they make
the choice themselves. Slightly exaggerating, mountaineers may be more afraid
during the flight to the Himalayas than when climbing Everest itself. Thus, ex-
tra safety has to be provided whenever drivers feel an automatic device takes
the process of determining what is risky out of their hands. This is a some-
what self-defeating process from the point of view of introducing an automated
environment.

8.4 Behavioural Adaptation

'Behavioural adaptation' is a summary descriptive term that stands for a number
of phenomena that may occur as a consequence of drivers interacting with the
newly introduced element in their task environment. The general connotation of
the concept is that it is detrimental to the positive effects originally foreseen to
result from the new support system. Two forms may be distinguished, direct and
higher-order behavioural adaptation.

8.4.1 Direct Changes in Behaviour

It has been established that drivers do indeed show riskier behaviours in several
important cases in which risk-attenuating devices were provided. One of these is
ABS, which was studied in the famous Munich taxi driver experiment (Aschen-
brenner et al., 1994). This study had both a retrospective and a prospective part,
and in neither could it be shown that ABS-vehicles were involved in less accidents
than standard vehicles. The other case is the one involving seat belts (Janssen,
1994a). Table 8.1 shows the results of an instrumented-vehicle study with subjects
138 Janssen

TABLE 8.1. Increase in average driving speed on motorway and in amount of

car-following at very short headways, over three consecutive measurements at about
4-month interval in first year of belted driving, compared to previously unbelted driving
(baseline) .
1st 2nd 3rd

Increase in driving speed (km/h; baseline: 112.3 kmlh) +2.2 +2.6 + 2.8
Increase in occurrence of following headways < 0.5 s +0.5 + 6.3 +7.9
(%; baseline: 5.5 %)

who originally were non-wearers of the belt and who became wearers for the pur-
pose of the study. This long-term study could apply an almost ideal experimental
design in that it had an alternating own vehicle/instrumented vehicle design; that
is, subjects started wearing the belt all the time in their own vehicle and came to
the laboratory for measurements in an instrumented vehicle at regular (4-months)
intervals after an initial beltless baseline measurement. Their driving performance
was compared to a control group of non-wearers who remained so for at least the
duration of the study.
While ABS and seat belts are apparently demonstrations of the existence of
the phenomenon, it is not yet clear (a) whether this will always happens or what
would distinguish cases in which they do from cases in which they do not and (b)
whether the compensation is complete and will totally eliminate the safety effect
that should follow from the engineering estimate.
To come to terms with these questions, we would need valid and quantitative
models of road user decision making. Elementary utility models (O'Neill, 1977;
Janssen and Tenkink, 1988) have already paid some services in this respect. In the
Janssen and Tenkink model (see Fig. 8.1), the road user is assumed to balance the
(dis)utilities of time loss during the trip, plus the possible accident risk, against
the utility of being at the destination. From this a choice of optimal speed, and
possibly of other driving behaviour parameters, then follows so as to be at the
optimum of that balance.
It has been derived, for example, from this type of consideration that a device
having an expected effectiveness (i.e. an engineering estimate) 8 will not reduce
accident risk with that same factor but with a factor that happens to be

== 1 - (1 - 8)-1/(c+l) , (1)

where c is Nilsson's (Nilsson, 1984) parameter in his speed-risk function, with

values between 3 and 7 for different types of accidents. For fatalities, c == 7.
It is clear that the safety effect to be realised will always be less than the ex-
pected effectiveness (see Fig. 8.2). For example, if we take the commonly used
value of 8 == 0.43 for the seat belt, the estimated effect to be achieved for the
fatality rate per km would be in the order of 7% rather than 43% (at 100%
use rate).
 8. Modelling Driver's Risk Taking Behaviour 139

occident loss, after

time loss
loss accident loss, befor~
per km

- - - total loss. otter

- ...
,
" total loss, before

o,....-~-------+---+------v

- resulting occident loss,

before & after
loss
per minute

occident loss, after

---------------+-- time loss

accident loss. before

FIGURE 8.1. Utility model (Janssen and Tenkink, 1988) shows how drivers select optimum
speed as a function of time (opportunity) losses and accident risk so as to make the resulting
total expected loss minimal. It appears to be generally true that whenever accident risk is
objectively reduced ('after' situation) the optimum speed that is selected will move towards
the higher end of the scale.

8.4.2 A Word of Caution About Working with Risk Measures

in Traffic Safety Studies
In the risk sciences, it is good to remind ourselves from time to time what is the
'risk' that we are dealing with. In traffic safety, for example, risk can be expressed
per kilometre travelled, per capita/vehicle (per year) or per unit of time spent in
140 Janssen

1.0 FIGURE 8.2. Expected and actual (i.e.

pre- or postdicted) safety benefit, ac-
cording to a simple utility model of
0.8 driver behaviour.
<W
en
(12
Q)
c 0.6
eu
>
.~

0
~
Q) 0.4
ii
:1
.....
U
co
0.2

0.0 0.2 0.4 0.6 0.8 1.0

effectiveness E of safety device

traffic. It should be realised that these indicators can go in different directions at

the same time and this has led to much confusion, e.g. in criticising Wilde's Risk
Homeostasis Theory or variations thereof. Likewise, the finding of some risk level
being not constant from one period of time to the next is not necessarily a valid
criticism of these formulations.

8.4.3 A Piece of Empirical Evidence from

Seat Belt Accident Statistics
Hardly does one ever have access to all components in the chain. However, there
appears to be one case in which almost all is available, which happens to be the
safety belt.
The statistics for this case comprise a set of data from the Federal Republic of
Germany pertaining to a sudden rise in seat belt wearing rate and its subsequent
effect on passenger car driver fatalities (Briihning et al., 1986).
From 1 August 1984, onward German road authorities exerted a strict en-
forcement of seat belt legislation by setting a fine of 40 Deutschmark ('Verwar-
nungsgeld') for being apprehended as a non-wearer passenger car driver. Almost
overnight, the overall wearing rate in the country rose from 58 to 92%. This makes
the German data as close as coming from an ideal 'natural' experiment as pos-
sible and it makes it feasible to postdict fatalities after the increase in wearing
rates, given the availability of an engineering estimate of the belt's effectiveness
in preventing a fatality when a crash happens.
Obviously, the spectacular increase in wearing rate should be reflected in an
immediate downward step in passenger car fatalities. As a matter of fact, depending
on the engineering estimate used for the seat belt's intrinsic effect the reduction
in fatalities should have been between 24% and 31% (that is, the increase in use
 8. Modelling Driver's Risk Taking Behaviour 141

rate multiplied by the engineering estimate). On the other hand the utility model,
on the basis of Equation (1), would have predicted a reduction of between 3% and
4%. The actual reduction was 6.7%. The readers may draw their own conclusions
as to which of the two was the best prediction (or rather, postdiction): Janssen,
(1994b), has further discussion 1.

8.4.4 Higher-Order Forms ofAdaptation

Following are the other forms of adaptation that may occur as the result of having
a new support available:

The generation of extra mobility (VMT). For example, navigation systems may
not so much reduce excess mileage as generate extra mileage into areas that
were formerly avoided. Or entrepreneurs who formerly 'lost' 5% or 6% of the
mileage driven by their fleet because drivers selected non-optimal routes to their
destination may now plan an extra trip a day because navigation performance
has become flawless.
Road use by less qualified segments ofthe driving population. It is to be expected
that some categories of users that did not dare to venture out in traffic, realis-
ing their own imperfections, will do so if offered an extra amount of 'built-in'
safety.
Driving under more difficult conditions. Similarly, the extra safety offered will
tempt road users to move to places they formerly avoided.

All these effects lend themselves to modelling by elementary utility considerations,

as discussed earlier.
On the level of aggregate accident statistics, a negative correlation between fatal
accident rates and VMT is predicted by utility considerations and has often been
observed. For example, a British investigation (Shannon, 1986) showed that for
the period from 1973 to 1983, there was a correlation of -0.88 between the annual
fatal accident rates (per mile) and VMT on British motorways. Exactly the same
value can be calculated from that paper for US Interstate data. Finally, in Japan
the death rate per km driven fell on average by a factor of 1.12 per year between
1966 and 1982, while the motorised mileage rose by an average of 1.08 from year
to year in that same period. The correlation between these two rates amounted
to -0.97. Thus, those years that were marked by relatively large decreases in the
death rate per km were also marked by relatively large increases in kilometre per
capita.
On a still higher level, many authors have noted that what is good for macro-
economy is bad for traffic safety and vice versa and that this is not just because of
economy-induced fluctuations in VMT (Joksch, 1984; Partyka, 1984; Wagenaar,
1984; Wilde, 1991). Table 8.2 lists correlations between annual unemployment

1 In which it is shown that selective recruitment is not an explanation of the less-than

expected safety effect in the German data.
142 Janssen

TABLE 8.2. Correlation between annual

unemployment rates and same-year per capita
traffic death rates (from Wilde, 1991).
United States, 1948-1987 -0.68
Sweden, 1962-1987 -0.69
West Germany, 1960-1983 -0.83
Finland, 1965-1983 -0.86
Canada, 1960-1986 -0.86
United Kingdom, 1960-1985 -0.88
The Netherlands, 1968-1986 -0.88

rates and same-year per capita traffic death rates in seven Western countries. Taken
together, the data do appear to confirm the basic assumption that safety and utility
are related to each other, i.e. safety is a factor in the utility considerations associated
with undertaking a trip.

8.5 The Link Between Behaviour and Risk

It would be of great help to the prediction of net safety effects if, given certain en-
gineering estimates, use rates, behavioural changes, etc., the relationships between
parameters conventionally used to describe driving behaviour and accident risk are
known. This would then permit the translation of effects observed in behavioural
studies in simulators and instrumented vehicles into safety effects to be expected
in the population. While the issues are still not resolved, progress has indeed been
made recently with respect to several important parameters.

8.5.1 Average Speed, Speed Variability and Risk

There presently exist several models that relate average speed and/or speed vari-
ability to accident risk. The most pragmatic and useful one is the one by Nilsson
(1984), who distinguishes between different types of accident (fatal; killed + se-
riously injured; material damage). The functions are power functions of average
driving speed, with powers ranging from three (material damage only) to seven
(fatal). With respect to speed variability and risk functions, useful and ready-to-use
expressions have been derived by Salusjarvi (1990).

8.5.2 Lane-Keeping Performance and Risk

The situation is somewhat less clear-cut for lateral performance indicators. It seems
that for the time being we have to be satisfied with proxies like TLC and the
frequency with which lane exceedances occur.
 8. Modelling Driver's Risk Taking Behaviour 143

8.5.3 Car-Following and Risk

In the seminal work of Farber (1993, 1994), a set of car-following data measured
in actual traffic was used to assess the impact of a collision-avoidance system
that would effectively reduce the following driver's reaction time to a sudden
braking action by the preceding vehicle. This can be generalised to calculating
the risk attached to a given situation per see The algorithm has the following steps
(Janssen, 2000):

For a given headway it is calculated whether, for a given range of reaction times,
a collision would follow if the preceding vehicle were to brake vehemently, i.e.
at full braking power.
The total probability of a collision is then computed by integration over a log-
normal distribution (which has a tail towards the longer reaction times) of driver
reaction times.
The mean and the standard deviation of the distribution are, moreover, adapted
to headway itself. This procedure was introduced by Farber so as to indicate that
drivers follow more attentively at shorter headways.
In case of a rear-end collision, the speed difference at the moment of impact is
computed. The overall risk of the car-following situation is then computed by
multiplying accident probability by the squared speed difference at impact.

Fig. 8.3 illustrates results for a few everyday car-following situations. As has
been observed by other authors, there is a 'worst' headway to follow, which is not
at the shortest range. This is intuitively clear when it is realised that although the
probability of the collision itself happening becomes higher at shorter headways,
its severity will be less because at a short headway the speed difference between
the two vehicles at the moment of impact will be less.

300 v1 =20 m/s

dec. =-8 m/s2
250

200

~ 150
a:
0

100

50

0
0.25 0.5 0.75 1 1.25 1.5 1.75 2
Headway

FIGURE 8.3. Rear-end collision risk when following a vehicle at a certain headway that
drives at 20 m/s and brakes suddenly (for two speeds of the following vehicle). Risk units
are arbitrary, i.e. defined as 100 at one of the configurations.
144 Janssen

8.6 Countermeasures Against Behavioural Adaptation

8.6.1 Should There Be Any?

The first thing to be kept in mind when pondering the question of what to do
against behavioural adaptation is that there is a good side to it as well. One has
better primary performance at a somewhat decreased accident risk. Likewise, the
extra mobility that is generated (grandmother gets out on the road again) is an
asset, i.e. a social gain.

8.6.2 Incentive Schemes and Their Expected Results

To move the optimal speed that drivers select, according to the utility model, in
one direction or the other, costs and benefits of driving in a more or less risky way
should be changed. This is the theoretical way out and empirical evidence from lab
studies and others indicates that this approach can work. The model then permits
to predict what safety effects to expect from these so-called incentive schemes
(e.g. Janssen, 1990).

8.7 Conclusions

There is now at least some useful knowledge available on the behavioural effects of
driver supports, which permits more than an educated guess on what safety effects
will follow from the provision of these supports. This derives both from a more
advanced insight into the behaviour itself and from the availability of procedures
to translate behavioural effects into safety effects. Thus

behavioural adaptation is definitely here to stay;

we better start modelling it, preferably in a quantitative way;
we should forget our prejudices against econometric principles, because people
are not mental cripples, as many (cognitive) psychologists want us to believe.

8.8 An Afterthought

When driver support systems that offer a safety benefit are introduced on the road
their benefits will, in all likelihood, be less than what originally could be expected.
This is, by itself, an important fact of life. However, as some authors have surmised,
the introduction of any specific safety measure could well be no more than a tiny
bubble on top of a continuously ongoing societalleaming curve of a much broader
nature. Although Smeed's ideas (Smeed, 1949, 1968, 1972; Smeed and Jeffcoate,
1970) about the learning process that comes along with a society's increasing level
of motorisation are no longer popular today, they may well reflect what is the more
significant permanent background against which all incidental safety measures
 8. Modelling Driver's Risk Taking Behaviour 145

become relatively minor. Modelling that background process is maybe really what
we should aim for in the long term.

References
Aschenbrenner, K., Biehl, B. and Wurm, G. (1994). Mehr Verkehrssicherheit durch bessere
Technik? Felduntersuchungen zur Risikokompensation am Beispiel des Antiblockiersys-
terns (ABS). Bast, Bergisch Gladbach, Bericht 8323.
Bruhning, E., Ernst, R., Glaeser, H.P., Hundhausen, G., Klockner, J.H. and Pfafferott, I.
(1986). Zum Riickgang der Getotetenzahlen im StraBenverkehr-Entwicklung in der
Bundesrepublik Deutschland von 1970 bis 1984. Zeitschrift fiir verkehrssicherheit, 32,
154-163.
Evans, L. (1985). Human behaviour feedback and traffic safety. Human Factors, 27, 555-
576.
Farber, E. (1993). Using freeway traffic data to estimate the effectiveness of rear-end colli-
sion countermeasures. Proceedings ofthe Third Annual Meeting ofthe Intelligent Vehicle
Society ofAmerica, Washington, DC.
Farber, E. (1994). Using the Reamacs model to compare the effectiveness of alternative rear
end collision warning algorithms. XIVth International Technical Conference on Enhanced
Safety of Vehicles, Miinchen. Paper Nr. 94 S3 0 03.
Janssen, W.H. (1988). Gurtanlegequoten und Kfz-Insassen-Sicherheit: eine Anmerkung zu
jiingsten deutschen Erkentnissen. Ze itsch rift fiir verkehrssicherheit. 34, 65-67.
Janssen, W.H. (1990). The economy of risk; or, what do I get for my error? Ergonomics,
33,1333-1348.
Janssen, W.H. (1994a). Seat-belt wearing and driving behaviour: An instrumented-vehicle
study. Accident Analysis and Prevention, 26, 249-261.
Janssen, W.H. (1994b). Behavioural adaptation to road safety measures: A framework and an
illustration. In R.M. Trimpop and GJ.S. Wilde (Eds.). Challenges to Accident Prevention.
Styx, Groningen, pp. 91-100.
Janssen, W.H. (2000). Functions relating driver behaviour and accident risk. Report TM-
00-D004. TNO Human Factors, Soesterberg, The Netherlands.
Janssen, W.H. (2001). Advanced driver supports, driver behaviour, and road safety. Pro-
ceedings ofthe 8th World Congress on Intelligent Transport Systems, Sydney. (CD-ROM)
Janssen, W.H., Kaptein, N. and Claessens, M. (1999). Behaviour and safety when driving
with in-vehicle devices that provide real-time traffic information. Proceedings of the 6th
World Congress on Intelligent Transport Systems, Toronto. (CD-ROM)
Janssen, W.H. and Tenkink, E. (1988). Considerations on speed selection and risk home-
ostasis in driving. Accident Analysis and Prevention, 20, 137-143.
Joksch, H.C. (1984). The relation between motor vehicle accidents deaths and economic
activity. Accident Analysis and Prevention, 16,207-210.
Nilsson ,G. (1984). Speeds, accident rates and personal injury consequences for different
road types. Report number 277VTI, Linkoping, Sweden.
O'Neill, B. (1977). A decision-theory model of danger compensation. Accident Analysis
and Prevention, 9, 157-165.
Partyka, S.C. (1984). Simple models of fatality trends using employment and population
data. Accident Analysis and Prevention, 16, 211-222.
Salusjarvi, M. (1990). Finland. In G. Nilsson (Ed.). Speed and Safety: Research Results
from the Nordic Countries. Linkoping.
146 Janssen

Shannon, H.S. (1986). Road-accident data: Interpreting the British experience with partic-
ular reference to the risk homeostasis theory. Ergonomics, 29(8), 1005-1015.
Smeed, R.I. (1949). Some statistical aspects of road safety research. Journal of the Royal
Statistical Society, Series A, Part I, 1-34.
Smeed, R.I. (1968). Variation in the pattern of accident rates in different countries and their
causes. Traffic Engineering and Control, 364-371.
Smeed, R.I. (1972). The usefulness of formulae in traffic engineering and road safety.
Accident Analysis and Prevention, 4, 303-312.
Smeed, R.I. and Jeffcoate, G.O. (1970). Effects of changes in motorisation in various
countries on the number of road fatalities. Traffic Engineering and Control, 150-151.
Wagenaar, A.C. (1984). Effects of macro-economic conditions on the incidence of motor
vehicle accidents.Accident Analysis and Prevention, 16, 191-205.
Wilde,G.I.S. (1991). Economics and accidents: A commentary. Journal of Applied Be-
haviour Analysis, 24, 81-84.
9
Dealing with Behavioural Adaptations
to Advanced Driver Support Systems
FARIDA SAAD

9.1 Introduction
Over the past 15 years major technological advances have been made in the field of
automotive technology. Many research and development programmes (in Europe,
Japan and the United States) have been devoted to the design and implementation
of new driver support systems and information management systems (for route
planning, obstacle detection, car-following situations, speed control and so on).
The development of these systems raises crucial issues at a technical level as
well as in terms of their consequences on driver activity (for an overview, see,
e.g., Michon, 1993; Parkes and Franzen, 1993; Noy, 1997). Some of these issues
deal, in particular, with the conditions of use of the systems, their effects on driver
behaviour and strategies and their impact on the operation and safety of the traffic
system. A major concern is about the 'behavioural adaptations' that may occur in
response to the introduction of these systems in the driving task and their impacts
on road safety (Smiley, 2000).
These new support systems will mediate drivers' interactions with their driving
environment (vehicle, road infrastructure, traffic rules, other road users) by creating
new sources of information and/or offering new modes of vehicle control. They
will thus alter the conditions in which the driving task is currently performed and
can thus be expected to engender changes in drivers' activity. Changes may occur
(1) within the very activity of 'supported' drivers (in terms of divided attention
between the new internal sources of information and direct monitoring of the road
environment, changes of driving strategies, delegation of control to the driving
support system and so on); (2) within their interactions with other road users (effect
on the behaviour of other road users, 'readability' of assisted drivers' behaviour for
other drivers, etc.). It is then important to specify the nature, direction and extent
of the changes likely to occur at these different levels, since they will determine
the ultimate impact on road safety (Evans, 1985; OCDE, 1990).
The changes associated with the use of these new support systems and their
acceptance by drivers will depend (1) on the types of task they are designed
to support (navigation, guidance or control tasks; Allen et al., 1971); (2) their
functions and the type of mediation they provide ('description' as regards the state

147
148 Saad

of the driving environment, 'prescription' as regards the regulating action the driver
has to take; 'intervention' or 'taking over' part of the driving task in the event of
deliberate driver delegation or of driver failure).
Up to now, most support systems are dedicated to specific driving tasks. Their
competence is by definition limited to the area of that task (or a subset of conditions
in which that task is performed, such as conditions of good visibility for instance).
The mediation offered is thus only partial, the driver's direct control over the
driving environment is always necessary and he remains responsible for the overall
management of his journey.
Studying the integration of the systems into the overall driving task and iden-
tifying behavioural changes when using them are thus critical aspects that need
to be carefully studied and analysed. This entails (Saad and Villame, 1999) the
following:
Taking account of the essential dimensions of the road environment in which
that activity takes place (nature of the interactions at work, regulatory, structural
and dynamic constraints, etc.). This reference to the context (Suchman, 1987)
is particularly important in view of the diversity and variability of the road
situations that drivers may encounter during a journey.
Choosing functional units of analysis making it possible to examine not only
the impact on the performance of the specific task to which the support system
is dedicated (compliance with safety margins or speed limits, for instance), but
also its compatibility with the performance of other driving tasks (overtaking,
interacting with other users, etc.).
Selecting the relevant indicators for revealing the changes likely to take place in
driver's activity.
These issues make direct demands on our knowledge of the driving task and of the
psychological processes (cognitive and motivational) that govern drivers' activity.

9.2 'Behavioural Adaptation' in Road Safety Research

In road safety research, the term 'behavioural adaptation' is mainly associated
with unintended or unexpected behavioural changes that may appear in response
to the introduction of a change in the traffic system and which may (more or less)
jeopardise road safety. Thus, the emphasis is placed primarily on the negative
aspects of the phenomenon.
For example, an OECD expert group (1990) defined behavioural adaptations
as 'those behaviours which may occur following the introduction of changes to
the road-vehicle-user system and which were not intended by the initiators of
the change'. On the basis of a review of a large number of empirical studies,
the expert group concluded that 'behavioural adaptation does occur, although not
consistently'. Behavioural adaptation may be an immediate response to the change
introduced in the traffic system or may only appear after a long time. Generally,
behavioural adaptation does not eliminate safety gains from measures, but tends
to reduce the size of the expected safety effects. Different elements are assumed
 9. Dealing with Behavioural Adaptations to Advanced Driver Support Systems 149

to influence the occurrence of behavioural adaptation such as the nature and the
'perceptibility' of the changes introduced in the traffic system (changes that directly
influence the way the driving task is performed or changes that alter the driver's
subjective safety, for instance), the degree of freedom that the change allows drivers
(changes that give the driver an opportunity for adapting his behaviour) or the
presence of competitive motives (safety versus mobility or productivity motives,
for instance).
Although behavioural adaptation is a widely acknowledged phenomenon, the
factors likely to explain it and the processes underlying its occurrence are not
clearly established. Numerous processes may in fact come into play between the
introduction of an 'innovation' in the traffic system and its 'adoption' by drivers,
its 'translation' into behaviour (whether 'safe' or 'risky') and its longer term con-
sequences on the operation and safety of the traffic system (Brown, 1985).
These processes should be analysed in greater depth, in particular those in-
fluencing the way drivers interact with their driving environment (vehicle, road
infrastructure, traffic rules, other road users, etc.). Such analyses should help to
formulate hypotheses about the changes in behaviour that may occur, identify the
conditions in which a 'negative' compensation for safety might appear and direct
thoughts on the means of minimising the extent of such negative changes.
Within the European project adaptive integrated driver-vehicle interface
(AIDE 1) , which aims to develop an harmonised interface that integrates all in-
vehicle support and information systems, a research activity is devoted specifi-
cally to identifying crucial behavioural adaptation issues associated with the use
of new support systems and determining the most relevant parameters that can be
implemented in models for supporting design and safety assessment processes.
The aim of this paper is to present the main phases of this research activity and
the results obtained so far. We begin by outlining the main results of a literature
review on 'behavioural adaptation' to new driver support systems (Saad et al.,
2004), especially advanced driver assistance systems (ADAS), which intervene
more or less directly in the performance of the driving task, such as adaptive cruise
control (ACC) or intelligent speed adaptation (ISA). We then describe the ongoing
activity associated with the conduct of experimental studies in order to improve
our knowledge on short and long term behavioural adaptation (Saad et aI., 2005)
and to develop models that can act as a reference for the design of an adaptive,
integrated in-vehicle interface supporting the multiple tasks drivers have to perform
in modem vehicles (Cacciabue and Hollnagel, 2005).

9.3 Behavioural Adaptation to Advanced Driver

Support Systems
In most research on 'behavioural adaptation' to new support systems, one of the ma-
jor concerns is the identification of 'adverse behavioural consequences' (Grayson,

1 AIDE is an EC Funded Project of the 6th Framework Programme - Project N. IST-l-

507674-IP
150 Saad

1996) or negative side effects associated with their use and which may reduce the
expected (safety) benefits of the assistance provided to the driver. In many research
studies, it is assumed that driving with systems which take over some elements
of the driving task (such as speed and time headway control) may reduce drivers'
workload and provide them with an opportunity for devoting less attention to the
(primary) driving task. Another concern relates to the drivers' ability to cope with
the limitation of support systems and to resume control in 'safety critical' traffic
scenarios, either because of drivers' misconceptions about the functioning of the
system and over-reliance on the system or as a consequence of drivers' reduced
attention to the driving task. In some studies, particular emphasis is placed on the
possible deterioration of drivers' interactions with other road users.
These research orientations guide the choice of the driving performance indica-
tors taken into account for assessing behavioural changes (such as driving speed,
safety margins in car-following situations and lateral control of the vehicle, per-
formance to a secondary task or subjective assessment of workload). They also
guide the choice of the driving situations or scenarios examined ('normal' driving
situations, 'safety critical' driving situations such as take-over situations in which
the driver has to regain control of his vehicle because of the limitations themselves
of the systems or a technical failure).
The second major concern relates to drivers' opinions and their acceptance of
the assistance provided. In most research studies, perceived usefulness of and sat-
isfaction about the systems are assessed either through standardised questionnaires
(see, e.g., Van der Laan et al., 1997) or through drivers' verbal reports and in-depth
interviews. In some studies, drivers' acceptance of the support system is more pre-
cisely assessed through the very usage of the system (in terms of drivers' decision
to engage the systems in various situational contexts and the overall duration of
system engagement, for instance).
Several empirical studies have been carried out, focusing mainly on the impact
of individual support systems, such as collision avoidance systems (CAS), speed
limiters, ISA or ACC systems, either in the' controlled' context of driving simulator
or in the complexity of real driving situations. Most of these studies have been
short-term studies and 'the effects (of support systems) on traffic safety and driver
behaviour are still uncertain in many respects' (Nilsson et al., 2002), especially their
long-term effects. Nevertheless, some critical issues have already been identified.
These issues are presented and discussed below.

9.3.1 The Diversity ofBehavioural Changes Studied

and Observed
The first critical issue encountered when examining the impact of a given support
system concerns the diversity of behavioural changes studied and observed, as
well as the magnitude and direction of these changes.
For example, the main behavioural changes observed when studying the be-
havioural impact of ACC are changes in speed, in the safety margins (time headway
 9. Dealing with Behavioural Adaptations to Advanced Driver Support Systems 151

or time-to-collision) adopted in various car-following situations (such as steady

car-following, catching up a slower vehicle, etc.) and in the lateral control of the
vehicle, as well as changes in lane occupancy and in the frequency of lane change
manoeuvres. The results obtained are sometimes contradictory. For example, in
some studies, the average driving speed increased when using ACC (Ward et al.,
1995; Hoedemaeker and Brookhuis, 1998), whereas in others this was not the case
(Hogema and Janssen, 1996; Stanton et al., 1997; Tornes et al., 2002). The same
divergent tendencies were observed when it came to the frequency of unsafe safety
margins adopted in car-following situations and the lateral position of the vehi-
cle when using ACC. For example, Fancher et al. (1998) and Saad and Villame
(1996) observed that the use of ACC reduced the frequency of short-time head-
way, Stanton et al. (1997) and Hoedemaeker and Kopf (2001) found no differences
between manual and ACC driving and Ward et al. (1995) observed a tendency to
drive with shorter time headways with ACC. Sometimes the results are more con-
vergent. For example, when driving with ACC on motorways, one observes a
general tendency for drivers to spend more time in the left lane (Hoedemaeker and
Brookhuis, 1998; Nilsson, 1995; Saad and Villame, 1996; Tornros et al., 2002).
In 'safety critical' scenarios, which require the drivers to reduce their speed or
to brake, drivers generally react later and/or with reduced safety margins when
driving with ACC (Nilsson and Nabo, 1996; Hoedemaeker and Brookhuis, 1998;
Rudin-Brown and Parker, 2004).
The same divergent trends are observed when studying the behavioural impact
of ISA. For example, changes in car-following behaviour were observed when
driving with ISA and the direction and magnitude of these changes varied from
one study to another. In urban areas, Varhelyi and Makinen (2001) and Hjalmdahl
and Varhelyi (2004) found that, when driving with ISA, the time headway in car-
following situations increased, whereas Carsten and Fowkes (2000) observed an
increase in close following (time headway less than 1 s). On rural roads, the results
are more convergent and suggest that there is an increase in close following.
As regards the workload associated with the use of the systems, the results also
reveal some variations. Nilsson (1995) and Ward et al. (1995) found no difference
in subjective workload between usual driving and ACC driving, while Hoede-
maeker and Brookhuis (1998) and Tornros et al. (2002) found that the subjective
workload was lower in ACC driving. Stanton et al. (1997) and Rudin-Brown and
Parker (2004) found that a secondary task was better performed when driving with
ACC, while this was not the case in the study of Young and Stanton (2002). Some
studies suggest that workload depends on the characteristics of the systems stud-
ied and/or the task to be carried out (Nilsson and Nabo, 1996; Hoedemaker and
Kopf, 2002; Young and Stanton, 2004). Other studies suggest that workload also
depends on drivers' degree of familiarisation with the system (Kopf and Nirschl,
1997).
Thus, part of the observed diversity may be due to the functional characteristics
of the systems studied (for ACC, control algorithm, time headway targets avail-
able, deceleration level used, etc.) and to the degree of drivers' experience with
them.
152 Saad

More generally, the diversity of the results obtained raises questions about the
methods used, the type and number of variables selected for assessing the im-
pact of the system, and finally the (implicit or explicit) models governing their
choice. When examining the results, attention should be paid to these theoretical
and methodological issues. In particular, the context in which the studies have
been carried out (driving simulator, closed tracks or real driving situations) should
be specified as well as the various scenarios and driving tasks in which the be-
havioural changes have been identified. The diversity of the results obtained could
then be examined and discussed in the light of the characteristics of these vari-
ous contexts and scenarios. Such analysis is particularly relevant insofar as the
situational context plays an important role in the behavioural changes observed
when driving