algorithms

Article
HMM Adaptation for Improving a Human Activity
Recognition System
Rubn San-Segundo *, Juan M. Montero, Jos Moreno-Pimentel and Jos M. Pardo
Speech Technology Group, E.T.S.I. Telecomunicacin, Universidad Politecnica de Madrid, 28040 Madrid, Spain;
juancho@[Link] (J.M.M.); josemoreno@[Link] (J.M.-P.); pardo@[Link] (J.M.P.)
* Correspondence: [Link]@[Link]; Tel.: +34-915-495700 (ext. 4228)

Academic Editor: Olga C. Santos

Received: 16 June 2016; Accepted: 29 August 2016; Published: 2 September 2016

Abstract: When developing a fully automatic system for evaluating motor activities performed
by a person, it is necessary to segment and recognize the different activities in order to focus
the analysis. This process must be carried out by a Human Activity Recognition (HAR) system.
This paper proposes a user adaptation technique for improving a HAR system based on Hidden
Markov Models (HMMs). This system segments and recognizes six different physical activities
(walking, walking upstairs, walking downstairs, sitting, standing and lying down) using inertial
signals from a smartphone. The system is composed of a feature extractor for obtaining the most
relevant characteristics from the inertial signals, a module for training the six HMMs (one per activity),
and the last module for segmenting new activity sequences using these models. The user adaptation
technique consists of a Maximum A Posteriori (MAP) approach that adapts the activity HMMs to the
user, using some activity examples from this specific user. The main results on a public dataset have
reported a significant relative error rate reduction of more than 30%. In conclusion, adapting a HAR
system to the user who is performing the physical activities provides significant improvement in the
systems performance.

Keywords: user adaptation; human activity segmentation; HMMs; smartphone inertial sensors

1. Introduction
The research on multisensor networks has increased significantly in the last 10 years, defining the
Internet of Things (IoT) concept. These networks typically include cameras, indoor location systems
(ILS), microphones, wearable sensors, etc. Using the information from sensors, computer-based
systems can adapt their behaviors to the context conditions (increasing their intelligence) or they can
report important information to the user (difficult to obtain through other means). Thanks to the
increment of sensor neural networks, the number of possible research areas has also increased rapidly.
One of these areas is psycho-motor training where an automatic system senses a psychical activity
carried out by a person and provides feedback about the performance. When developing a fully
automatic system for evaluating motor activities, one important aspect is to segment and recognize the
different activities in order to focus the system analysis on some specific ones. This process must be
carried out by a Human Activity Recognition (HAR) system. The recognition of human activities has
received a lot of attention in the last five years due to the high number of promising applications and
the increasing interest shown by government and commercial organizations.
This paper proposes a user adaptation technique for improving a HAR system based on Hidden
Markov Models (HMMs). This system segments and recognizes six different physical activities
(walking, walking upstairs, walking downstairs, sitting, standing and lying down) using inertial
signals from a smartphone. This paper is organized as follows. Section 2 presents the state of the art.
Section 3 shows an overview of the HMMs-based HAR system, describing the main modules. Section 4

Algorithms 2016, 9, 60; doi:10.3390/a9030060 [Link]/journal/algorithms

Algorithms 2016, 9, 60 2 of 13

describes the user Maximum A Posteriori (MAP) adaptation. Section 5 presents the experiments
carried out in this work, including a detailed description of the dataset used in the experiments. The
main discussions and conclusions are summarized in Section 6.

2. Background
HAR systems can be categorized according to the sensor type or the machine learning algorithm.
According to the sensor type, it is possible to consider various types of sensors: on-body, object-placed
or ambient sensors. On example of environment sensors is video cameras in monitoring areas [1,2].
Human activity can be also analyzed based on a rich variety of acoustic events. The determination
of both the identity of sounds and their position in time may help to detect and describe that human
activity [3]. Most environment sensors require an important infrastructure support: for example, the
installation of video cameras in the monitoring areas. Additionally, people not always spend all their
time in the same environment. In this respect, environmental sensors are limited by their infrastructure
and cannot provide monitoring outside the specific environment. This limitation can be overtaken
using on-body sensors [4,5]. Body-worn sensors add new possibilities to the human monitoring
system [6]: not only by being able to measure body signals (e.g., physiological, motion, location) but
also by providing portable and off-site user supervision at any location without the need of fixed
infrastructure. In the literature there are different approaches for locating motion sensors in different
body parts such as the waist, wrist, chest and thighs achieving good classification performance [79].
In [10], a chest-mounted accelerometer was used for classifying five Activities of Daily Living (ADL).
However, the use of body sensors has important limitations such as the users discomfort while
wearing them (these sensors are usually uncomfortable for the common user) and energy-limited
mobile devices (they do not provide a long-term solution for activity monitoring).
In recent years, smartphones and smartwatches have become widespread, increasing the number
of possibilities for human-centered applications. These devices include embedded built-in sensors
such as microphones, dual cameras, accelerometers, gyroscopes, etc. The inertial sensors are a
very interesting for monitoring ADL. These devices have important advantages [11,12]: easy device
portability, unobtrusive sensing provided by the embedded sensors and the processing power of new
smartphones that allow online computation. Because of this, some works focused on HAR using
smartphones have been developed [1317].
HAR is a machine learning problem, where a system extracts features from sensor signals,
generates a model for each activity, and classifies the next activities based on these models. In the
literature, different machine learning solutions have been applied to the recognition of activities
including Naive Bayes [18], Decision Trees [19], and Support Vector Machines (SVMs) [20]. In many
works, several approaches are compared: for example, Yang [21] uses the WEKA learning toolkit
to compare the accuracy rates of several machine learning approaches: C4.5 Decision Trees, Nave
Bayes, k-Nearest Neighbor, and Support Vector Machines. In [13], three learning algorithms were
evaluated: Logistic Regression, J48, and Multilayer Perceptron. Hidden Markov Models is a successful
modeling strategy for classifying temporal sequences. HMMs offer dynamic time warping, have
clear Bayesian semantics and well-understood training algorithms. HMMs are very robust against
degradation, giving the possibility to be trained on one person and to be tested on another. In the last
five years, there has been an increase in the number of HAR systems based on HMMs for modeling
inertial signals (Table 1).
Algorithms 2016, 9, 60 3 of 13

Table 1. Comparison of previous HAR works based on HMMs.

Ref. Sensor Target Classes Users Time User Adaptation Performance

Five bi-axial accelerometers,
Activity recognition
Mannini and Sabatini [11] located at the hip, wrist, 7 activities 13 29 min NO Error: 1.6%
and segmentation
arm, ankle and thigh
Accelerometer in a
Lee and Cho [22] Activity recognition 7 activities 3 339 min NO Error: 15.0%
LG smartphone
Transportation mode
Reddy et al. [23] Accelerometer in a Nokia n95 5 modes 16 1200 min NO Precision and Recall > 93%
recognition
Tri-axial accelerometer
Wang et al. [24] Activity recognition 6 activities 13 ~100 min NO Error: 2.8%
(MMA7260)
2 sedentary
Witowski et al. [25] Simulated data Detecting physical activity behavior vs. - 1000 days NO Error: 18.8%
physical activity
Three MTx 3-DOF (Degree of Activity recognition
Trabelsi et al. [26] 12 activities 6 ~100 min NO Error: 9.6%
Freedom) inertial trackers and segmentation
Accelerometer in a Activity recognition Error: 3.5% (without
San-Segundo et al. [17] 6 activities 30 220 min NO
Samsung Galaxy S2 and segmentation sequence model)
Accelerometer in a Activity recognition and
This paper 6 activities 30 220 min YES Error: 2.0%
Samsung Galaxy S2 segmentation
literature there are some references proposing other user adaptation techniques: a semi-supervised
method [27] or a Multi-Classifier Adaptive-Training (MCAT) algorithm [28]. The MCAT algorithm
consists of using a meta-classifier for combining several pattern recognition methods.
This meta-classifier is trained with user-dependent adaptation data to improve the results for a
Algorithms 2016, 9, 60
specific user. 4 of 13

3. HAR System Overview

According to Table 1, this work is the first one (known by the authors) that proposes a user
Figuretechnique
adaptation 1 shows for theimproving
general system architecture
the performance of apresented
HMMs-based in [17],
HARincluding
system. Inantheadditional
literature
module for user MAP adaptation. The system is made up of four main modules:
there are some references proposing other user adaptation techniques: a semi-supervised method [27] feature extraction,
HMMs
or training, HMMs
a Multi-Classifier adaptation and
Adaptive-Training activity
(MCAT) recognition/segmentation.
algorithm [28]. The MCAT algorithm The main contributions
consists of using
aofmeta-classifier
this paper are for
focused on HMMs
combining training
several patternand adaptationmethods.
recognition [Link] In order to considerisonly
meta-classifier the
trained
influence
with of the HMMs
user-dependent in the results,
adaptation data totheimprove
Activitythe
Sequence Model
results for (ASM)user.
a specific proposed in [17] has been
deactivated in this work.
3. HAR The System
feature Overview
extraction module obtains the accelerometer and gyroscope signals, samples them
withFigure
a sampling ratethe
1 shows of general
50 Hz and filters
system them for noise
architecture reduction
presented in [17],(with a 20 Hz
including an cut-off frequency).
additional module
This sampling rate is sufficient for capturing human body motion:
for user MAP adaptation. The system is made up of four main modules: feature extraction,more than 95% of its energy
HMMs is
contained below 15 Hz [5]. Using a Butterworth low-pass filter (with a 0.3
training, HMMs adaptation and activity recognition/segmentation. The main contributions of thisHz cut-off frequency), the
sensorare
paper acceleration
focused onsignals
HMMsare divided
training andinto body acceleration
adaptation [Link] gravity.
In order The Euclidean
to consider only the magnitude
influence
and time derivatives (jerk da/dt and angular acceleration dw/dt) are also obtained
of the HMMs in the results, the Activity Sequence Model (ASM) proposed in [17] has been deactivated during the feature
extraction
in this [Link].

Figure 1.
Figure 1. HAR
HAR system
system architecture.
architecture.

The sample sequences are grouped together in frames: fixed-width sliding windows of 2.56 s
The feature extraction module obtains the accelerometer and gyroscope signals, samples them
and 50% overlap (128 samples per frame with an overlap of 64 samples). From each frame, the
with a sampling rate of 50 Hz and filters them for noise reduction (with a 20 Hz cut-off frequency).
system obtains a feature vector computing measures from the inertial signals. These features are
This sampling rate is sufficient for capturing human body motion: more than 95% of its energy is
traditional measures such as the mean, correlation, signal magnitude area (SMA) and auto
contained below 15 Hz [5]. Using a Butterworth low-pass filter (with a 0.3 Hz cut-off frequency),
regression coefficients [29]. This vector has been extended including more features from time and
the sensor acceleration signals are divided into body acceleration and gravity. The Euclidean magnitude
frequency domains generating a vector with a total of 561 features. The dataset used in this work
and time derivatives (jerk da/dt and angular acceleration dw/dt) are also obtained during the feature
(available at the UCI Machine Learning Repository) includes the already-computed feature vectors.
extraction process.
For comparison, this work uses the same features proposed in [30]. The set of features computed
The sample sequences are grouped together in frames: fixed-width sliding windows of 2.56 s
from the time domain are mean, standard deviation, median absolute deviation, max, min,
and 50% overlap (128 samples per frame with an overlap of 64 samples). From each frame, the
magnitude, energy, interquartile range, entropy, autoregression coefficients with the Burg order
system obtains a feature vector computing measures from the inertial signals. These features are
equal to 4, and the correlation coefficient between different axis. The set of features obtained from
traditional measures such as the mean, correlation, signal magnitude area (SMA) and auto regression
the frequency domain includes additional ones such as the index of the frequency component with
coefficients [29]. This vector has been extended including more features from time and frequency
domains generating a vector with a total of 561 features. The dataset used in this work (available at the
UCI Machine Learning Repository) includes the already-computed feature vectors. For comparison,
this work uses the same features proposed in [30]. The set of features computed from the time domain
are mean, standard deviation, median absolute deviation, max, min, magnitude, energy, interquartile
range, entropy, autoregression coefficients with the Burg order equal to 4, and the correlation coefficient
between different axis. The set of features obtained from the frequency domain includes additional
ones such as the index of the frequency component with the largest magnitude, the weighted average
Algorithms 2016, 9, 60 5 of 13
Algorithms 2016, 9, 60 5 of 13

the largest
of the magnitude,
frequency the weighted
components, skewness,average of the
kurtosis andfrequency
the energy components, skewness,
in 64 bins covering thekurtosis and
whole signal
the energy in
frequency range. 64 bins covering the whole signal frequency range.
For
For HMMs development,we
HMMs development, weused
used thethe
HTKHTK toolkit
toolkit [31].[31].
ThisThis
toolkittoolkit
allows allows us tothe
us to train train
HMMs,the
HMMs, adapt them to a new user, and recognize new activities using
adapt them to a new user, and recognize new activities using these HMMs. In the HAR used in these HMMs. In the HAR used
in this
this work,six
work, sixHMMs
HMMsare areconsidered,
considered,one onefor
forevery
everyactivity.
activity. Every
Every model
model represents
represents the the sequence
sequence
of observed feature vectors corresponding to each activity. Given a vector
of observed feature vectors corresponding to each activity. Given a vector sequence, it is compared sequence, it is compared
to
to all
all the
the models,
models, computing
computing their their likelihood.
likelihood. The Themodel
modelwith withthe
thehighest
highestlikelihood
likelihoodisisthe the activity
activity
recognized.
recognized. A HMM can be seen as a finite state machine in which all the states can change every
A HMM can be seen as a finite state machine in which all the states can change every
time
timeunit
[Link]
Everystate
statejjgenerates
generatesaafeature
featurevector
vectorOtOtconsidering
consideringaaprobability
probability density
density bj(Ot).
bj(Ot). There
There
isisalso
alsoaa probability
probability toto control
control transitions
transitions between
between states.
states. For Forexample,
example,the thetransition
transitionbetween
betweenstatestateii
aa state
state jj can
can be
be governed
governed by by the
the discrete
discrete probability
probability aaijij. .Figure
Figure22represents
representsaaHMM HMMwith withsix sixstates,
states,
associated
associated with a sequence of six observations O1 to O6. In HTK, the entry and exit states ofaaHMM
with a sequence of six observations O1 to O6. In HTK, the entry and exit states of HMM
are
are non-emitting.
[Link] Thisisisto
tofacilitate
facilitatethe theconstruction
constructionof ofcomposite
compositemodelsmodels(out (outofofthe
thescope
scopeof of this
this
paper). The association between the observed vector and states is X =
paper). The association between the observed vector and states is X = 1; 2; 3; 3; 4; 5; 5; 6. 1; 2; 3; 3; 4; 5; 5; 6.

Markov
a22 a33 a44 a55
Model
a12 a23 a34 a45 a56
1 2 3 4 5 6
b3(O2) b3(O3) b5(O6)

Observation
Sequence

O1 O2 O3 O4 O5 O6

Figure
[Link]
Exampleof
ofaaHidden
HiddenMarkov
MarkovModel.
Model.

The
The joint
joint probability
probability that O is
that O is generated
generated by
by the
themodel
modelMMmoving
movingthrough
throughthe
thestate
statesequence
sequenceXXis
iscalculated
calculatedsimply
simply as the product of the transition probabilities and the output probabilities:
as the product of the transition probabilities and the output probabilities:
P(PO , XX/M
(O, = aa12 bb2 ((O
/ M)) = )a b (O )a b (O )...
O1 ) a 23 b 3(O 2) a 33b 3(O )3 ... (1)(1)
12 2 1 23 3 2 33 3 3

In a real application, only the observation sequence O is known and the underlying state
In a real application, only the observation sequence O is known and the underlying state sequence
sequence X is hidden. In this case, the required likelihood is computed by considering only the most
X is hidden. In this case, the required likelihood is computed by considering only the most likely state
likely state sequence. The main parameters of every model are aij and bj(Ot). The output distribution
sequence. The main parameters of every model are aij and bj (Ot ). The output distribution bj (Ot ) can be
bj(Ot) can be modeled by using Gaussian mixtures, reducing the model parameters to the mean and
modeled by using Gaussian mixtures, reducing the model parameters to the mean and variance of
variance of every Gaussian distribution. The model parameters can be determined automatically
every Gaussian distribution. The model parameters can be determined automatically through a robust
through a robust and efficient re-estimation procedure (estimation maximization, EM) considering a
and efficient re-estimation procedure (estimation maximization, EM) considering a set of training
set of training examples corresponding to a particular activity:
examples corresponding to a particular activity:
(1) Initialize all Gaussian distributions with the mean and variance computed throughout the
(1) Initialize all Gaussian distributions with the mean and variance computed throughout the
whole dataset.
whole dataset.
(2) Calculate the forward and backward probabilities for all states j and times t.
(2) Calculate the forward and backward probabilities for all states j and times t.
(3) For each state j and time t, use the probability Lj(t) and the current observation vector Ot to
(3) For each state j and time t, use the probability Lj(t) and the current observation vector Ot to
update the accumulators for that state.
update the accumulators for that state.
(4) Use the final accumulator values to calculate new parameter values.
(5) If the the
(4) Use final
value of accumulator
P = P(O/M) forvalues to calculate
this iteration new
is not parameter
higher values.
than the value at the previous iteration,
(5) then
If thestop; of P = P(O/M)
valueotherwise, repeatforthe
this iteration is not steps
aforementioned higherusing
than the
the value at the previous
new re-estimated iteration,
parameter
then stop; otherwise,
values (from step 2). repeat the aforementioned steps using the new re-estimated parameter
values (from step 2).
In the estimation maximization (EM) algorithm, the main target is to maximize the likelihood of
In therespect
an activity estimation
to itsmaximization
HMM: Maximum (EM) Likelihood
algorithm, the main target
Estimation (MLE).is toInmaximize the likelihood
some applications, it is
of an activity
possible respectatodiscriminative
to consider its HMM: Maximum trainingLikelihood
[Link] (MLE). Inthe
In this procedure, some applications,
main target is toit
maximize the differences between models. The HTK toolkit includes a tool for training HMMs in a
Algorithms 2016, 9, 60 6 of 13
Algorithms 2016, 9, 60 6 of 13

discriminative
is way. In
possible to consider this work, the
a discriminative Maximum
training Mutual
procedure. Information
In this Estimation
procedure, the (MMIE)
main target is to
discriminative
maximize procedurebetween
the differences has alsomodels.
been evaluated [32]. Inincludes
The HTK toolkit this procedure,
a tool for the mainHMMs
training difference
in a
compared to MLE is the objective function to be optimized. In the case of the MMIE, the function
discriminative way. In this work, the Maximum Mutual Information Estimation (MMIE) discriminative to
maximize has
procedure is: also been evaluated [32]. In this procedure, the main difference compared to MLE is the
objective function to be optimized. In the case of the MMIE, the function to maximize is:
P(Oi , Xi / Mi )
FMMIE = log

FMMIE =i = activities log P(OPi , (XOi /M
i
, Xi )n / Mi ) (2)
(2)
i = activities
n = activities sequences P(Oi , Xn /Mi )
n= activities sequences
The first term in the numerator is identical to the objective function for the MLE. In order to
maximizefirst
The term in(2),
Equation the the
numerator
numeratoris identical
must betoincreased
the objective
whilefunction for the MLE.
the denominator is In order to
decreased.
maximize
Similar to Equation
the MLE,(2),thethe numerator
MMIE has themust betoincreased
target maximize while the denominator
the likelihood of eachisobservation
decreased. Similar
given byto
the MLE, the MMIE has the target to maximize the likelihood of each observation
training sequences. In addition, the MMIE has a denominator term that can be reduced by given by the training
sequences.
decreasing theIn addition, the MMIE
probabilities of otherhas a denominator
possible term that In
activity sequences. canconclusion,
be reducedthe byMMIE
decreasing the
attempts
probabilities of other
make the correct possible activity
hypothesis [Link]
more probable, In conclusion,
at the same the time,
MMIEitattempts
tries to make
makethe correct
incorrect
hypothesis
hypothesesmore probable, and at the same time, it tries to make incorrect hypotheses less probable.
less probable.
Continuous
Continuousactivity
activityrecognition
recognition andand
segmentation
segmentation involves connecting
involves several HMMs
connecting several inHMMs
[Link]
In the HTKIntoolkit,
sequence. the HTKthe Viterbi
toolkit,algorithm is expanded
the Viterbi algorithmtoisallow severaltomodels
expanded allow to be connected
several models in to the
be
search space:
connected the search
in the last state of every
space: model
the last is of
state connected
every modelwith istheconnected
first statewith
of allthe
models (Figure
first state 3).
of all
Each
modelsmodel in the
(Figure 3). sequence
Each model corresponds directly
in the sequence to its activity.
corresponds directly to its activity.

a22 a33 a44 a55 a22 a33 a44 a55 a22 a33 a44 a55

a12 a23 a34 a45 a56 a12 a23 a34 a45 a56 a12 a23 a34 a45 a56
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6

walking walking walking

a22 a33 a44 a55 a22 a33 a44 a55 a22 a33 a44 a55

a12 a23 a34 a45 a56 a12 a23 a34 a45 a56 a12 a23 a34 a45 a56
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6

walking-upstairs walking-upstairs walking-upstairs

a22 a33 a44 a55 a22 a33 a44 a55 a22 a33 a44 a55

a12 a23 a34 a45 a56 a12 a23 a34 a45 a56 a12 a23 a34 a45 a56
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6

walking-downstairs walking-downstairs walking-downstairs

a22 a33 a44 a55 a22 a33 a44 a55 a22 a33 a44 a55

a12 a23 a34 a45 a56 a12 a23 a34 a45 a56 a12 a23 a34 a45 a56
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6

standing standing standing

a22 a33 a44 a55 a22 a33 a44 a55 a22 a33 a44 a55

a12 a23 a34 a45 a56 a12 a23 a34 a45 a56 a12 a23 a34 a45 a56
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6

sitting sitting sitting

a22 a33 a44 a55 a22 a33 a44 a55 a22 a33 a44 a55

a12 a23 a34 a45 a56 a12 a23 a34 a45 a56 a12 a23 a34 a45 a56
1 2 3 4 5 6 1 2 3 4 5 6 1 2 3 4 5 6

laying laying laying

Figure 3.
Figure 3. Search space for
Search space for activity
activity segmentation.
segmentation.

4. User
4. UserMaximum
MaximumA
APosteriori
PosterioriAdaptation
Adaptation
In order
In order to
to generate
generate activity
activity HMMs
HMMs adapted
adapted toto every
every user,
user, the
the system
system trains
trains general
general activity
activity
HMMs including
HMMs including information
information from
from all
all users.
users. After training these
After training these general
general models,
models, user-adapted
user-adapted
activity HMMs are generated by adapting the general models to every user
activity HMMs are generated by adapting the general models to every user via a Maximum via a Maximum A A
Posteriori approach (sometimes referred to as Bayesian adaptation). MAP adaptation
Posteriori approach (sometimes referred to as Bayesian adaptation). MAP adaptation needs prior needs prior
knowledge about
knowledge aboutthethe model
model parameter
parameter distribution
distribution (original
(original HMMs). HMMs).
For MAPFor MAP adaptation
adaptation purposes,
purposes, the informative priors that are generally used are the user-independent
the informative priors that are generally used are the user-independent model parameters. The updated model
parameters. The updated formula for the parameter
formula for the parameter of mixture component m is: of mixture component m is:

NNmm MU
mm =
= + MU USER
m +
generalm m
USERm (3)
(3)
Nmm ++ MU
N MU general
N + MU
Nmm + MU
The estimated m is a linear combination of the parameter of mixture component m in the
original model (generalm) and the parameter of mixture component m obtained considering only
Algorithms 2016, 9, 60 7 of 13

The estimated m is a linear combination of the parameter of mixture component m in the

original model (generalm ) and the parameter of mixture component m obtained considering only
observation vectors from the user (USERm ). MU is the adaptation coefficient that defines the weighting
of user-dependent information compared to general HMMs. Nm is the occupation likelihood of the
adaptation data (from the user) along the T frames, defined as:

t= T
Nm = N (Ot ; generalm , ) (4)
t =1 general m

After training the adapted activity HMMs for every user, the activity recognition and segmentation
process is the same as the method used with user-independent HMMs: the process consists of
computing the likelihood of the best model sequence when generating several activities.

5. Experiments Carried out in This Work

This section describes the dataset used in the experiments, the baseline results considering the
training algorithms (MLE and MMIE), the user adaptation experiments, and at the end, the final results
with a discussion.

5.1. Dataset Used in the Experiments

This work has been carried out using a public dataset available at the UCI Machine Learning
Repository: the Human Activity Recognition Using Smartphones Data Set [30]. This dataset contains
inertial information (from smartphone sensors: accelerometer and gyroscope) recorded from a group
of 30 people (from 19 to 48 years old), performing six different physical activities several times.
These activities are walking, walking upstairs, walking downstairs, sitting, standing and lying down.
While performing these activities, every user carried a smartphone (Samsung Galaxy S II) for recording
the inertial signals. These signals consisted of the three-axial linear acceleration and the three-axial
angular velocity being sampled at a constant rate of 50 Hz. The dataset contains 13,182 s of recording
including 400 activity instances from 30 users. An example of the recording process can be seen in a
video [33].
In this work, the main aim is to recognize the activity sequence carried by every user. For this
evaluation, all activities carried out from the same user have been stored in the same file, defining
a recording session. There are 30 sessions with an average number of 13.3 activities per session.
The problem to solve in this paper consists of recognizing and segmenting all physical activities
recorded in the same session (Figure 4). In the initial configuration of the dataset, it was divided in
two sets, and 70% of the users were selected for training and 30% for testing the system. In this work,
the 30 sessions have been randomly divided into six subsets. Every session includes all activities
carried out by the same user, so all activities from the same user are included in the same subset. This
characteristic avoids the person-dependent characteristic being influenced by the activity recognition or
segmentation. In order to improve the significance of the results, a six-fold cross-validation procedure
has been carried out. The cross-validation procedure uses four subsets for training the activity Hidden
Markov Models (HMMs), one for validation (tuning the system parameters) and one for testing. This
configuration has been repeated six times in a round-robin strategy. The results presented in this paper
are average values obtained throughout the six-fold cross-validation procedure. In every experiment,
the system is evaluated with all sessions (13,182 s) (defining a 95% confidence interval of 0.4%).
In Sections 5.2 and 5.3, only the validation results were considered for tuning the adaptation coefficient
and selecting the training procedure. The final results are presented in Section 5.4 with the testing
subsets, using the best system configuration.
Algorithms 2016, 9, 60 8 of 13
Algorithms
Algorithms2016,
2016,9,9,60
60 88of
of13
13

Figure
Figure4.
Figure [Link]
4. Exampleof
Example ofactivity
of activityrecognition
activity recognitionand
recognition andsegmentation.
and segmentation.
segmentation.

Regarding
Regarding the
the evaluation
evaluation metrics,
metrics, in
in this
this work
work we
we considered
considered the
the Activity
Activity Recognition
Recognition Error
Error
Regarding the evaluation metrics, in this work we considered the Activity Recognition Error Rate
Rate
Rate(ARER):
(ARER):the
thepercentage
percentageofoftime
timethat
thathas
hasbeen
beenwrongly
wronglyassigned
assignedto
toan
anactivity
activity(Equation
(Equation(5)).
(5)).
(ARER): the percentage of time that has been wrongly assigned to an activity (Equation (5)).
Time
Time(sec)
(sec)wrongly
wronglyclassified
classified
ARER
ARER(%)
(%) ==100
100 Time (sec) wrongly classi f ied (5)
(5)
ARER(%) = 100 Session (5)
Sessionduration
Session duration ((sec)
duration (sec)
sec)
In
In addition
In addition to
addition tothis
to thismeasure,
this measure,other
measure, otherpossible
other possiblemetrics
possible metrics
metricsare
are precision
areprecision
precision and
and
andrecall.
recall. The precision
TheThe
recall. precision isis the
precision the
is
time
time correctly
correctly assigned
assigned to
to this
this activity
activity (true
(true positive)
positive) divided
divided by
by the
the
the time correctly assigned to this activity (true positive) divided by the activity time detected byactivity
activity time
time detected
detected
by
by the
the the system
system system (including
true true
(including
(including and
truefalse
and false
false positive
and positive times).times).
positive Recall Recall
times). Recall
is isis defined
defined defined as
as the
as the time the time
time correctly
correctly correctly
assigned
assigned
assigned
(true (true
(truepositive)
positive) positive)
divided by divided
divided by
bythe
the actual the actual
actualduration
duration duration of
ofthe
theactivity.
of the activity. activity.
As it was As
Asitshown
itwas
wasshown
shown
in [17], in [17],
[17],there
inthere there
is an
is
is an
an important
important correlation
correlation between
between these
these three
three measures.
measures. Because
Because ofof this,
this,
important correlation between these three measures. Because of this, ARER will be considered for ARER
ARER will
will be
be considered
considered
for
for system
systemsystem development,
development,
development, andand
and
in thein the
the final
infinal final experiments,
experiments,
experiments, all
all metrics
metrics
all metrics will
will bewill be
be provided
providedprovided for
for comparison
for comparisoncomparison with
with
with further
further
further works. works.
works.

5.2.
[Link]
BaselineExperiments
Baseline ExperimentsConsidering
Experiments ConsideringDifferent
DifferentTraining
TrainingAlgorithms
Algorithms
Figure
Figure 55 represents
Figure represents the ARER
the ARER depending
ARER depending
depending on on the
on the training
the training procedure.
procedure. As
training procedure. As itit is
is shown,
is shown, the the
discriminative training
discriminativetraining strategy
trainingstrategy (MMIE)
strategy(MMIE) obtains
(MMIE)obtains
obtains slightly
slightly better
better results,
results, although
although the
the differences
differences
slightly better results, although the differences are not are
are
not
notstatistically
statistically
statistically significant.
significant.
significant. For the
thenext
Fornext
For the next experiments,
experiments,
experiments, the
theMMIE
the MMIE MMIE strategy
strategystrategy has
hasbeen
has been been considered.
considered.
considered. In
In this this
Intask,
this
task,
task,
the the
the system
system is ableisis
system toable
able to
to distinguish
distinguish
distinguish between between
between
static andstatic
static and
and dynamic
dynamic dynamic activities,
activities,activities, but
but the
but the confusionthe confusion
confusion
between
between
between
static static
static activities
activities activities
(standing, (standing,
(standing,
sitting and sitting
sitting
lying and
and lying
lyingisdown)
down) down) isis very
very high. veryA high.
high.
similarA
A similar
similar behavior
behavior behavior occurs
occurs for occurs
the
for
for the
the dynamic
dynamic activitiesactivities
dynamic activities
of walking,of
of walking, walking
walking,upstairs
walking walkingand upstairs
upstairs
walking and
and walking
walking downstairs.
downstairs. downstairs.
Figure 5 also Figure
Figure
includes55 also
also
the
includes
includes the
the error
error points
points for
for all
all the
the subjects
subjects using
using the
the representation
representation method
method
error points for all the subjects using the representation method proposed in [34]. Regarding the ARER proposed
proposed in
in [34].
[34].
Regarding
Regardingthe
distributions, ARER
theboth
ARER distributions,
distributions,
training procedures both
both training
veryprocedures
training
show procedures show
showvery
similar behaviors. verysimilar
similarbehaviors.
behaviors.

Activity
ActivityRecognition
Recognition Error
ErrorRate
Ratevs.
vs. Training
Trainingprocedure
procedure
55 55
(%)
Rate (%)
Error Rate

44 44
Recognition Error

33 33
Activity Recognition

22 22

11 11
Activity

00 00
MLE
MLE MMIE
MMIE
Training
Training procedure
procedure

Figure
Figure 5.
Figure [Link]
Activity recognition
Activityrecognition error
recognitionerror rate
errorrate depending
ratedepending
dependingonon
on the
thethe training
training
training procedure
procedure
procedure including
including
including data
data
data points
points
points for
for
forall
allthe
the subjects
subjectsusing
usingthe
therepresentation
representation method
method proposed
proposed
all the subjects using the representation method proposed in [34]. in
in[34].
[34].
Algorithms 2016, 9, 60 9 of 13
Algorithms 2016, 9, 60 9 of 13

Algorithms 2016, 9, 60 9 of 13
5.3. User Adaptation Experiments
[Link]
Userthe Adaptation Experiments
user adaptation experiments, all user sessions used for testing the system are randomly
divided For into
into
thetwo twosub-sessions
user sub-sessions
adaptation including
including
experiments, 50% 50%
all of sessions
ofuser
the the activities
activities each.
[Link]
used The Thethe
average
testing average
duration
system duration
areofrandomlyof an
an activity
activity
isdivided
440 s. The is 440
into s.
first
two The first
sub-session sub-session
sub-sessions is used is used
for adapting
including for
50% ofthe adapting
theHMMs the
to the
activities HMMs
userThe
each. to the
and averageuser
second and the second
sub-session
duration of anis
sub-session
used for testing
activity is used
is 440 theThe
s. for testing
system.
first the system.
There is notisany
sub-session There
used is adapting
overlap
for not any overlap
between thethese
HMMsbetween
two thethese
usertwo
sub-sessions.
to andItsub-sessions.
is
theimportant
second
It isremark
important
tosub-session that isto remark
the
used that the
sub-session
for testing sub-session
used
the usedisisfor
for testing
system. There testing
the
not same
any isalong
the same
overlap along
all the
between all the
twoexperiments
experiments
these to allowtoa
sub-sessions.
allow a fair
It iscomparison.
fair comparison.
important to remark that the sub-session used for testing is the same along all the experiments to
allow Figure
a fair66comparison.
represents
representsthe theevolution
evolutionof of thetheARERARER depending
depending on theon adaptation
the adaptation coefficient. This
coefficient.
representation
This representation also includes
Figure 6 represents the the
also includes confidence
evolution of theintervals
the confidence ARER at 95%
depending
intervals along
at 95% thethe
on along curve.
adaptation
the When
curve. MUMU
coefficient.
When =This
0 =no0
adaptation
norepresentation
adaptation is done
is alsoand
done the the
includes
and system
the obtains
confidence
system the same
obtains intervals
the sameresults
at as when
95%
results along
as using
when general
theusing
curve. HMMs.
When
general MU
HMMs. When 0 MU
= When
no
increases,
MU adaptation
increases,theis ARER
done
the ARER decreases
and the
decreasesuntil
system reaching
obtains
until a minimum
the same
reaching results as
a minimum forforMU
when = =15.
MUusing [Link]
Afterthis
general value,
HMMs.
this value, the
the ARER
When MU
ARER
increases,
increases the ARER decreases until reaching a minimum for MU =
increases because the available user-dependent data is limited and it is not productive to increase its
because the available user-dependent data is limited and it 15.
is After
not this value,
productive to the ARER
increase its
increases
weight
weight because
for HMMs
for HMMs training. the available user-dependent data is limited and it is not productive to increase its
training.
weight for HMMs training.
Activity Recognition Error Rate vs. Adaptation coefficient
4 Activity Recognition Error Rate vs. Adaptation coefficient
(%)(%)

4
Rate
Rate

3
Error

3
RecognitionError

2
Recognition

1
Activity

1
Activity

0
00 5 10 15 20 25 30
0 5 10 15 20 25 30
Adaptation Coefficient
Adaptation Coefficient
Figure 6. Activity recognition error depending on the adaptation coefficient.
[Link]
Figure Activity recognition
recognition error
error depending on the
depending on the adaptation
adaptationcoefficient.
coefficient.

The next
Thenext figure
nextfigure (Figure
figure(Figure 7)
(Figure7) shows
7)shows
showsthethe ARER
theARER depending
ARERdepending
dependingon on the
on the amount
amountof ofdata
datatotoadapt.
[Link]
Instead
The the amount of data to adapt. Instead of
ofofusing
using the whole
the whole sub-session
sub-session (50% of the user session) for adapting the HMMs, different amounts
using the whole sub-session (50%(50% of user
of the the user session)
session) for adapting
for adapting the HMMs,
the HMMs, different
different amounts
amounts have
have been
beenconsidered:
haveconsidered:
considered: 10%, 20%,
10%,30%,
20%,40%30%, 40%
30%,and and
40%50%
andof50%
50% of the
the user
user session.
session.
been 10%, 20%, theofuser session.
Activity
ActivityRecognition
Recognition Error
Error Rate vs. Amount
Rate vs. Amountof
ofdata
data
44
(%)
Rate(%)
ErrorRate

33
Recognition Error
Activity Recognition

22

11
Activity

00
0%0% 10%
10% 20%
20% 30%
30% 40%
40% 50%
50%
Amount of
Amount of data
data (%
(% of
of the
the user
user session)
session)
Figure7.
Figure
Figure [Link]
Activityrecognition
Activity recognition
recognition error
error
error rate
rate
rate depending
on on
depending
depending thethe
on the amount
amount
amount of
of data
of datadata for
forHMMs
for HMMsHMMs adaptation
adaptation
adaptation (% of
(%(% of the user session).
theof thesession).
user user session).
Algorithms 2016, 9, 60 10 of 13

As it is shown in Figure 7, the ARER decreases when we increase the amount of data
(this representation also includes the confidence intervals at 95% along the curve). Regarding the
differences between users, the absolute ARER reduction varies from 0.5% (the lowest reduction) to
1.5% (the highest reduction). The curve presented in Figure 7 does not show any saturation tendency,
so it means that if we increase the adaptation data, the ARER can continue decreasing. For further
works, a bigger dataset will be considered in order to analyze this effect.
In order to complete the analysis, a new experiment has been carried by training the HAR system
with the data of a single user. In this case, the first user sub-session has been used for training the
system (instead of adapting the system) and the second sub-session for testing. In this case, the ARER
increases to 9.4%. This significant degradation is due to the important reduction in the amount of data
for training the system. This result supports the utility of the adaptation algorithm proposed in this
paper as the best solution for developing a user-dependent HAR system when there is a small amount
of data per user.

5.4. Final Experiments and Discussion

This subsection presents the final results on test datasets. These experiments have been carried
out considering the best system configuration obtained from the analyses done on the validation sets
(see previous subsection): MMIE HMMs training and user MAP adaptation using 50% of the user
session for adaptation (first sub-session) and 50% for testing (second sub-session). The sub-session
used for testing is the same in all the experiments. It is important to remark that there is not any
overlap between testing and adaptation data. Table 2 shows the final results obtained on validation
and test subsets. This table includes the activity segmentation error rates (%), recall (%), precision (%)
and confidence intervals at 95%. The results on validation subsets have already been presented in
previous subsections.

Table 2. Final segmentation results including activity recognition error rate (ARER), recall and
precision metrics.

Validation Test
User Adaptation
ARER (%) ARER (%) Recall (%) Precision (%)
Baseline 3.5% 0.4% 3.5% 0.4% 92.9% 0.4% 92.6% 0.4%
MMIE training 3.3% 0.4% 3.3% 0.4% 93.2% 0.4% 93.1% 0.4%
MMIE training + User MAP adaptation 2.0% 0.3% 2.1% 0.3% 95.2% 0.3% 95.1% 0.3%

As shown, the final results on test subsets are slightly worse because the system was optimized
on the validation subset. Similar to the conclusion obtained in previous subsections, the ARER error
decreases significantly when adapting the HMMs to the user considering a MAP approach.
Table 3 includes the same experiments considering the original dataset partition (70% for HMMs
training and 30% for testing) for comparing to previous works. These results show that using a
user MAP adaptation, it is possible to significantly improve the segmentation results obtained in
previous works on this dataset. When adapting the HMMs to the user, the HAR has better modeling
for recognizing the physical activities carried out by this specific user. This adaptation is the main
contribution of this paper.
Algorithms 2016, 9, 60 11 of 13

Table 3. Final results considering the original dataset partition including activity recognition error rate
(ARER), recall and precision metrics.

Test
System
ARER (%) Recall (%) Precision (%)
Anguita et al. 2013 [30] 4.0% 0.4% - -
San-Segundo et al. 2016 [17] 3.2% 0.3% 93.3% 0.4% 93.1% 0.4%
MMIE training 3.1% 0.3% 93.8% 0.3% 93.9% 0.3%
MMIE training + user MAP adaptation 2.0% 0.3% 95.3% 0.3% 95.2% 0.3%

6. Conclusions
This work has proposed a user adaptation technique for improving a HAR system based on
HMMs. This system segments and recognizes six different physical activities (walking, walking
upstairs, walking downstairs, sitting, standing and lying down) using inertial signals from a
smartphone. The system is composed of a feature extractor for obtaining the most relevant
characteristics from the inertial signals, a module for training the six HMMs (one per activity), and
the last module for segmenting new activity sequences using these models. This paper has evaluated
two different HMMs training strategies: the first one, a generative approach (Maximum Likelihood
Estimation), and the second, a discriminative one (Maximum Mutual Information Estimation).
The discriminative training strategy (MMIE) obtains slightly better results, although the differences
are not statistically significant.
The main contribution of this paper is the analysis of a user adaptation technique for adapting
the HMMs to the user who performs the different activities. The user adaptation technique consists
of a Maximum A Posteriori (MAP) approach. The final results on a public dataset [30] have reported
significant error rate reduction: from 3.2% to 2.0% ARER (more than 30% relative error rate reduction).
In conclusion, adapting a HAR system to the user who is performing the physical activities reports
significant improvement in the activity segmentation process.

Acknowledgments: This work has been supported by ASLP-MULAN (TIN2014-54288-C4-1-R) and NAVEGABLE
(MICINN, DPI2014-53525-C3-2-R) projects.
Author Contributions: Rubn San-Segundo implemented the HAR system using HMMs, executed the
experimental work for the adaptation process, analyzed the results, drafted the initial manuscript and revised the
manuscript. Juan Manuel Montero conceptualized the HMM adaptation process, supervised the experiments and
revised the manuscript. Jos Moreno-Pimentel integrated the discriminative training procedure, executed the
experiments related to discriminative training, helped to draft the initial manuscript and revised the final version.
Jos Manuel Pardo analyzed the results, provided feedback and revised the final manuscript.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Poppe, R. Vision-based human motion analysis: An overview. Comput. Vis. Image Underst. 2007, 108, 418.
[CrossRef]
2. Poppe, R. A survey on vision-based human action recognition. Image Vis. Comput. 2010, 28, 976990.
[CrossRef]
3. Temko, A. Acoustic Event Detection and Classification. Ph.D. Thesis, Polytechnic University of Catalonia,
Barcelona, Spain, 2009.
4. Lukowicz, P.; Ward, J.A.; Junker, H.; Stger, M.; Trster, G.; Atrash, A.; Starner, T. Recognizing workshop
activity using body worn microphones and accelerometers. In Proceedings of the 2nd International
Conference Pervasive Computing, Vienna, Austria, 2123 April 2004; pp. 1822.
5. Karantonis, D.M.; Narayanan, M.R.; Mathie, M.; Lovell, N.H.; Celler, B.G. Implementation of a real-time
human movement classifier using a triaxial accelerometer for ambulatory monitoring. IEEE Trans. Inf.
Technol. Biomed. 2006, 10, 156167. [CrossRef] [PubMed]
Algorithms 2016, 9, 60 12 of 13

6. Bao, L.; Intille, S.S. Activity Recognition from User-Annotated Acceleration Data; Kanade, T., Kittler, J.,
Kleinberg, J.M., Mattern, F., Mitchell, J.C., Nierstrasz, O., Rangan, C.P., Steffen, B., Terzopoulos, D.,
Tygar, D., et al., Eds.; Pervasive Computing: Linz/Vienna, Austria, 2004; pp. 117.
7. Casale, P.; Pujol, O.; Radeva, P. Human activity recognition from accelerometer data using a wearable device.
In Pattern Recognition and Image Analysis; Springer: Berlin/Heidelberg, Germany, 2011; p. 289.
8. Krishnan, N.; Narayanan, C.; Colbry, D.; Juillard, C.; Panchanathan, S. Real time human activity recognition
using tri-axial accelerometers. In Proceedings of the Sensors, Signals and Information Processing Workshop,
Sedona, AZ, USA, 1114 May 2008.
9. Nishkam, R.; Nikhil, D.; Preetham, M.; Littman, M.L. Activity recognition from accelerometer data.
In Proceedings of the Seventeenth Conference on Innovative Applications of Artificial Intelligence, Pittsburgh,
PA, USA, 913 July 2005; pp. 15411546.
10. Hanai, Y.; Nishimura, J.; Kuroda, T. Haar-like filtering for human activity recognition using 3d accelerometer.
In Proceedings of the IEEE 13th Digital Signal Processing Workshop and 5th IEEE Signal Processing
Education Workshop (DSP/SPE), Marco Island, FL, USA, 47 January 2009; pp. 675678.
11. Mannini, A.; Sabatini, A.M. Machine learning methods for classifying human physical activity from on-body
accelerometers. Sensors 2010, 10, 11541175. [CrossRef] [PubMed]
12. Vinh, L.T.; Lee, S.; Le, H.X.; Ngo, H.Q.; Kim, H.I.; Han, M.; Lee, Y.-K. Semi-markov conditional random fields
for accelerometer-based activity recognition. Appl. Intell. 2011, 35, 226241. [CrossRef]
13. Kwapisz, J.R.; Weiss, G.M.; Moore, S.A. Activity recognition using cell phone accelerometers.
SIGKDD Explor. Newslett. 2011, 12, 7482. [CrossRef]
14. Brezmes, T.; Gorricho, J.L.; Cotrina, J. Activity recognition from accelerometer data on a mobile phone.
In Distributed Computing, Artificial Intelligence, Bioinformatics, Soft Computing, and Ambient Assisted Living;
Springer: Berlin/Heidelberg, Germany, 2009; pp. 796799.
15. San-Segundo, R.; Montero, J.M.; Barra-Chicote, R.; Fernndez, F.; Pardo, J.M. Feature Extraction from
Smartphone Inertial Signals for Human Activity Segmentation. Signal Proc. 2016, 120, 359372. [CrossRef]
16. Wu, W.; Dasgupta, S.; Ramirez, E.E.; Peterson, C.; Norman, G.J. Classification accuracies of physical activities
using smartphone motion sensors. J. Med. Intern. Res. 2012, 14, e130. [CrossRef] [PubMed]
17. San-Segundo, R.; Lorenzo-Trueba, J.; Martnez-Gonzlez, B.; Pardo, J.M. Segmenting human activities based
on HMMs using smartphone inertial sensors. Pervasive Mob. Comput. 2016, 30, 8496. [CrossRef]
18. Jatoba, L.C.; Grossmann, U.; Kunze, C.; Ottenbacher, J.; Stork, W. Context-aware mobile health monitoring:
Evaluation of different pattern recognition methods for classification of physical activity. In Proceedings
of the 30th Annual International Conference of the IEEE Engineering in Medicine and Biology Society,
Vancouver, BC, USA, 2025 August 2008.
19. Maurer, U.; Smailagic, A.; Siewiorek, D.; Deisher, M. Activity recognition and monitoring using multiple
sensors on different body positions. In Proceedings of the International Workshop on Wearable and
Implantable Body Sensor Networks (BSN06), Cambridge, MA, USA, 35 April 2006.
20. Anguita, D.; Ghio, A.; Oneto, L.; Parra, X.; Reyes-Ortiz, J.L. Energy Efficient Smartphone-Based Activity
Recognition using Fixed-Point Arithmetic. J. Univ. Comput. Sci. 2013, 19, 13951314.
21. Yang, J. Toward physical activity diary: motion recognition using simple acceleration features with mobile
phones. In Proceedings of the 1st ACM International Workshop on Interactive Multimedia for Consumer
Electronics (IMCE 09), Beijing, China, 23 October 2009.
22. Lee, Y.S.; Cho, S.B. Activity Recognition Using Hierarchical Hidden Markov Models on a Smartphone with
3D Accelerometer. In Proceedings of the 6th International Conference, Wroclaw, Poland, 2325 May 2011;
Volume 6678, pp. 460467.
23. Reddy, S.; Mun, M.; Burke, J.; Estrin, D.; Hansen, M.; Srivastava, M. Using mobile phones to determine
transportation modes. ACM Trans. Sens. Netw. 2010, 6, 13. [CrossRef]
24. Wang, J.; Chen, R.; Sun, X.; She, M.F.H.; Wub, Y. Recognizing Human Daily Activities from Accelerometer
Signal. Procedia Eng. 2011, 15, 17801786. [CrossRef]
25. Witowski, V.; Foraita, R.; Pitsiladis, Y.; Pigeot, I.; Wirsik, N. Using Hidden Markov Models to Improve
Quantifying Physical Activity in Accelerometer DataA Simulation Study. PLoS ONE 2014, 9, e114089.
[CrossRef] [PubMed]
Algorithms 2016, 9, 60 13 of 13

26. Trabelsi, D.; Mohammed, S.; Chamroukhi, F.; Oukhellou, L.; Amirat, Y. An unsupervised approach for
automatic activity recognition based on hidden Markova model regression. IEEE Trans. Autom. Sci. Eng.
2013, 3, 829335. [CrossRef]
27. Cvetkovic, B.; Lutrek, M.; Kalua, B.; Gams, M. Semi-supervised Learning for Adaptation of Human Activity
Recognition Classifier to the User. In Proceedings of the Twenty-Second International Joint Conference on
Artificial Intelligence (IJCAI11), Barcelona, Spain, 1622 July 2011; pp. 2429.
28. Cvetkovic, B.; Kalua, B.; Gams, M.; Lutrek, M. Adapting activity recognition to a person with
Multi-Classifier Adaptive Training. J. Ambient Intell. Smart Environm. 2015, 7, 171185.
29. Khan, A.M.; Lee, Y.-K.; Lee, S.Y.; Kim, T.-S. Human activity recognition via an accelerometer
enabled-smartphone using kernel discriminant analysis. In Proceedings of the 5th International Conference
on Future Information Technology, Busan, Korea, 2123 May 2010; pp. 16.
30. Anguita, D.; Ghio, A.; Oneto, L.; Parra, X.; Reyes-Ortiz, J.L. A Public Domain Dataset for Human Activity
Recognition Using Smartphones. In Proceedings of the 21th European Symposium on Artificial Neural
Networks, Computational Intelligence and Machine Learning (ESANN), Bruges, Belgium, 2426 April 2013.
31. Young, S.; Evermann, G.; Gales, M.J.F.; Hain, T.; Kershaw, D.; Liu, X.; Moore, G.; Odell, J.; Ollason, D.;
Povey, D.; et al. The HTK Book; Cambridge University Engineering Department: Cambridge, UK, 2006.
32. Chow, Y.L. Maximum mutual information estimation of HMM parameters for continuous speech recognition
using the N-best algorithm. In Proceedings of the International Conference on Acoustics, Speech, and Signal
Processing (ICASSP-90), Albuquerque, NM, USA, 36 April 1990; Volume 2, pp. 701704.
33. Activity Recognition Experiment Using Smartphone Sensors. Available online: [Link]
watch?v=XOEN9W05_4A (accessed on 30 August 2016).
34. Weissgerber, T.L.; Milic, N.M.; Winham, S.J.; Garovic, V.D. Beyond bar and line graphs: Time for a new data
presentation paradigm. PLoS Biol. 2015, 13. [CrossRef] [PubMed]

2016 by the authors; licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC-BY) license ([Link]