2011 6th International ICST Conference on Communications and Networking in China (CHINACOM)

Road Condition Monitoring Using On-board Three-

axis Accelerometer and GPS Sensor

Kongyang Chen1, Mingming Lu1,2, Xiaopeng Fan1, Mingming Wei1, and Jinwu Wu3
1
Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences，Shenzhen, 518055, China
2
Central South University, Changsha, 410083, China
3
Shenzhen Transportation Depot Group, Shenzhen, 518040, China
Email :{[Link], lumm, [Link], [Link]}@[Link], wujw@[Link]

Abstract—A study by US Federal Highway Administration has considered to be disqualified. In China, the road condition level
shown that road condition is an essential factor of highway is classified into four levels, according to Technical Code of
quality and smooth roads will lead to more comfortable driving Maintenance for Urban Road CJJ36-2006, which is one of the
experience and less municipal investment. International most important industry standards [20]. Current equipment
Roughness Index (IRI) has been widely used to measure used in measuring road condition is surveyed in [4], which is
pavement smoothness because it can provide a consistent rating composed by accelerometers, distance measuring instruments,
for different measurement tools. However, existing measuring graphic displays or something other instrument. But these road
tools based on IRI are usually very expensive. In this paper, we condition systems are much too expensive, which cost 8,000 to
present a low-cost vehicle-based solution, Road Condition
220,000 dollars. Therefore, these methods are too costly to be
Monitoring with Three-axis Accelerometers and GPS Sensors
(RCM-TAGPS), by using a cheap three-axis accelerometer and a
adopted in the whole urban district.
GPS sensor embedded in a vehicle to monitor the road condition. Different statistic metrics: Pavement roughness [5] is one
We analyze the Power Spectral Density (PSD) of pavement of the most important performance metrics of road condition,
roughness, estimate IRI, and classify the pavement roughness and it can evaluate whether the driving is comfortable and
level into four levels according to a Chinese industry standard. safety or not. Power Spectral Density (PSD) is the frequency
Experimental results show that RCM-TAGPS can evaluate characteristic of pavement roughness in the vehicle industry
pavement roughness level correctly, even under some and academia [6]. Moreover, International Roughness Index
interference like potholes, manholes and decelerating belts, and
(IRI) is another important metric, which was firstly adopted by
the total cost of RCM-TAGPS in each vehicle is no more than 50
dollars, which is about 1/4400 to 1/160 of the existing system used
the World Bank in the 1980s, and then widely used for
in civil engineering and municipal engineering. evaluating the pavement roughness [7-8]. Therefore, one of our
objectives in this paper is to find out the relationship between
Keywords-Road condition monitoring; three-axis acclerate sensor; PSD and IRI, in order to evaluate pavement roughness
Power Spectral Density(PSD); International Roughness Index(IRI); comprehensively.
pavement roughness level. This paper describes the design, implementation and
experimental evaluation of a road condition monitoring system,
I. INTRODUCTION namely Road Condition Monitoring with Three-axis
Road condition monitoring [1-3] is a challenging problem Accelerometer and GPS Sensor (RCM-TAGPS). Each vehicle
in the field of road transportation infrastructure all over the is embedded with a three-axis accelerometer and a GPS sensor
world. A zone in bad condition may damage the vehicles on it, when traveling around the road to be monitored. Each vehicle
endanger the drivers, and even cause accidents. City also collects the information of three-axis acceleration and
municipalities have spent millions of dollars to maintain and velocity with a high frequency. We then estimate PSD and IRI
repair these roadways. However, it is still a challenging of the pavement roughness, based on these data on
problem to keep these roadways in good condition. accelerations and velocity. Finally, we classify the actual road
condition into four different levels according to Technical Code
Expensive road monitoring system: City municipalities of Maintenance for Urban Road CJJ36-2006; in order to
are much concerned with global roughness information of a provide useful suggestions for road maintenance.
certain road, because the whole road would be repaired if it is
In section II, we provide an overview of the system
architecture of RCM-TAGPS. In section III, we describe the
This work was partially supported by the National Natural Science
Foundation of China under Grant No.60903222, the National S&T details of data collection by three-axis accelerometers and GPS
Major Project of China under Grant No.2009ZX03006-001-01, sensors. In section IV, we discuss a novel pavement roughness
Guangdong S&T Major Project under Grant No.2009A080207002 level algorithm based on PSD and IRI. We present the
and the National High Technology Research and Development evaluation of RCM-TAGPS in section V. Finally, we conclude
Program of China (863 Program) under Grant No. 2011AA010500 this paper in section VI.

978-1-4577-0101-6/11/$26.00 ©2011 IEEE

1032 978-1-4577-0101-6/11/$26.00 © 2011 IEEE

 II. SYSTEM ARCHITECTURE Our GPS receiver is the vehicle-mounted HM-CZ02 GPS
RCM-TAGPS consists of a set of sensor embedded vehicles sensor. The communication baud of this GPS module is 2400
for data collection and a central server for evaluating the bits per second, while the GPS data is updated once per second.
pavement roughness levels, as illustrated in Figure 1. RCM-TAGPS reads accurate time, location and velocity if we
follow the GPRMC format [10].
The client microprocessor is a 32 bits high performance
Cortex ARM chip STM32F103. It is with a high clock
frequency up to 72MHz, which is sufficient for our system. It
can generate electronic signals to control time sequence of
RCM-TAGPS, read the three-axis accelerometer and GPS
sensor, and send data to the central server after processed.
Simultaneously, we enable the build-in hardware watchdog,
monitoring the real-time statues of software operations to
ensure high reliability of the whole system.
Figure 1. System architecture

Raw sensing data are collected by accelerometers and GPS

devices. In RCM-TAGPS, we collect three direction
accelerations from the accelerometer embedded in a traveling
vehicle, and the current time, the velocity and the location from
GPS sensors.
Due to the complicated environment and the precision of
sensor devices, there are some noisy data. We apply the outlier
removal method for data cleaning, and the interpolate points in
curves where the data are missing. Each item in the set of the
processed sensing data is a 4-tuple as follows:
< time, location, velocity, three-axis acceleration >
Sensor-equipped vehicles upload these processed data to
the central server which maintains a database of all the detected
data. The central server analyzes PSDs and IRIs of the
pavement surface roughness, resulting in the comprehensive Figure 2. System hardware design
pavement roughness level according to the Technical Code of
Maintenance for Urban Road CJJ36-2006, one of industry
standard of the People's Republic of China [20]. B. Data cleaning algorithm
As we mentioned above, RCM-TAGPS reads 100 samples
III. DATA COLLECTION from the traveling vehicle’s three directions accelerations and
one sample of GPS data (including time, location and velocity)
RCM-TAGPS should provide continuous acceleration data per second, which is called raw sensing data.
from a three-axis accelerometer, and accurate time, location
and velocity from a GPS sensor. In this section, we will During our experiments, we identify several technical
describe our system design and data cleaning algorithm. challenges. Firstly, it is possible to miss some GPS data
because the GPS sensor may not work properly in urban
A. System hardware design canyons among tall buildings and tunnels [11]. Secondly, the
GPS receiver may not work well and provide valid GPS data
In RCM-TAGPS, the data collection module is composed
for about several seconds at the beginning of the starting.
of a three-axis accelerometer [9], a GPS sensor and a client
Thirdly, a transmission error may occur when transmitting a
microprocessor. Our system hardware design is shown in
large amount of data composed of accelerations and GPS data.
Figure 2.
Finally, the collected data from sensors are not normalized and
The high performance MEMS motion module, LIS33DE, is the units used in GPS do not follow international unit.
adopted as the three-axis accelerometer. It is with ± 2g To deal with these problems, we design the following data
dynamically measurement scale and 10000g high shock cleaning algorithm. We firstly delete the “bad zone” without
survivability. This module is with IIC/SPI digital output GPS data, including the first several seconds after a GPS sensor
interface, 400kHz read/write clock frequency and 100Hz powers on. Then, we remove the data when there is a
sampling frequency, so that it can be easily integrated into transmission error or a miss of GPS data, and record the index
various kinds of host devices. Additionally, the power at the missing points. After the removal of outlier data, we
consumption of this sensor is less than 1mW, which results in insert some interpolated points in curves where there is an
power budget saving for the whole mobile embedded system. outage. Our proposed algorithm generates these interpolation
points at an interval of 10ms to keep the continuity of all the

1033
collected data. Finally, we normalize these sensor data to the ⎧ 1 ∞
⎪ S (ω ) = 2π ∫−∞ R(τ )e
− jωτ
international system of units. Specifically, we normalize the dτ
three- axis accelerations to the range from -2g to 2g, where g ⎨ (1)
equals to 9.8 m/s2, and change the unit of the velocity from sea ⎪ R(τ ) = ∞ S (ω )e jωτ d ω
mile per hour to kilometer per hour by using the conversion ⎩ ∫−∞
factor 1.852. where τ represents the time interval, ω represents the angular
Our data cleaning algorithm results in the following frequency, S (ω ) is the PSD roughness metric, and R (τ ) is
sampling information: <time, location, velocity, three-axis temporal auto-correlation function. R (τ ) can be calculated by
acceleration>. The continuous three-axis acceleration is shown
the following equation:
in Figure 3, where the curves are X-axis acceleration, Y-axis
acceleration and Z-axis acceleration against sampling points R(τ ) = E[ x(t ) x(t + τ )] (2)
from the top to the bottom in the Figure.
where E[i] represents the expectation of the random process
that can be estimated from the following equation:
1 τ

t →∞ τ ∫0
E[ x(t )] = lim x(t )dt (3)

As demonstrated by Eq. 1, S (ω ) is the bilateral Power

Spectral Density (BPSD). However, only the single Power
Spectral Density (SPSD) is used in the engineering projects.
Therefore, we change the spectral density from the negative
frequency to the positive frequency as follows

⎧2 S (ω ) f ≥0
S ' (ω ) = ⎨ (4)
⎩0 f <0

Figure 3. The continuous three-axis accelerations where S ' (ω ) is the SPSD of pavement roughness.

Specially, if τ = 0 , we can deduce Eq.5 from Eqs.1~4 as

IV. PAVEMENT ROUGHNESS LEVEL ALGORITHM follows:
In this section, we present our pavement roughness level ∞ ∞
algorithm. The definition and solution of Power Spectral R(0) = E[ x 2 (t )] = ∫ S (ω ) d ω = ∫ S ' (ω ) d ω (5)
−∞ 0
Density (PSD) are described in Part A. Then we introduce the
International Roughness Index (IRI) and describe how to
calculated IRI by PSD in Part B. According to the Chinese B. International Roughness Inedex (IRI)
industry standard, the classification for pavement roughness International Roughness Index (IRI), adopted by the World
levels is presented as a function of IRI in Part C. Finally, we Bank in the 1980s, is the most widely used metric that is used
propose the details of the pavement roughness level algorithm to evaluate pavement roughness in highway transportation
in Part D. agencies. IRI is defined as the responses of the quarter car
model [17-18] at a speed of 80km/h, where the vehicle
A. Power Spectral Density (PSD) parameters are recommended by the highway safety research
institute (HDRI).
Power Spectral Density (PSD) is one of the most important
metrics of pavement roughness, which is wildly used in the However, it is still difficult to realize the definition and the
vehicle industry and academia. Compared with other metrics, quarter car model. As discussed above, PSD can easily be
PSD demonstrates the frequency characteristic of pavement estimated by the Fourier transformation of the temporal auto-
roughness, reflects the amplitude of roughness, and also shows correlation function. Hence, we aim to find out the relationship
the excitation energy under frequency variation [12-13]. between IRI and PSD, so that we can obtain IRI easier. Here,
the roughness standard deviation is considered as the
Experimental results show that pavement roughness can be
intermediate variable.
modeled as a Gaussian distribution with zero mean and
isotropic random character around the vibration source in the Firstly, we analyze the relationship of PSD and the
spatial domain. Based on the stochastic process theory, roughness standard deviation. Based on the probability and
pavement roughness can be described as a normal stationary statistics theories, the relationship between the standard
random process in the time domain. deviation and the expectation can be described as follows:
According to the Wiener-Khintchine theorem [14-16], the
σ 2 = E 2 [ x(t ) ] + E ⎡⎣ x 2 (t ) ⎤⎦ (6)
following equations are composed of a couple of Fourier
transform:

1034
 As the vehicle vibration can be modeled as a Gaussian D. Algorithm framework
distribution with zero mean and isotropic random character The framework of our pavement roughness level algorithm
[12], we can infer that the expectation is approximate to zero. is illustrated in Figure 4.
Hence, the standard deviation can be simplified as follows:
The first step is data collection, including the data from
σ 2 = E ⎡⎣ x 2 (t ) ⎤⎦ (7) three-axis accelerometers and GPS sensors. The second step is
carried out by the PSD analyzer. The auto-correlation function
Combined with Eq.5, the relationship between the standard will produce the PSD. The third step is carried out by the IRI
deviation and SPSD are described as follows: analyzer. Roughness standard deviation and IRI are calculated.
The fourth step is the classification of pavement roughness
∞ levels. We calculate RQI and classify pavement roughness into
σ= ∫0 S ' (ω )d ω (8)
four levels.
Secondly, we analyze the relationship between IRI and
roughness standard deviation.
As we know, contrast experiments on pavements at Beijing,
Changchun, and Nanjing have been carried out by the Ministry
of Communications Highway Scientific Research Institute
since 1990s [19]. The regression analysis results show that the
relationship between IRI and the roughness standard deviation
can be approximated by the following regression equation: Figure 4. Pavement roughness level algorithm framework
σ − 0.013
IRI = (9) V. EVALUATION
0.5926
This section presents the evaluation of RCM-TAGPS. The
experiment setting, including the road condition and the
C. Pavement roughness level classification
placement of our sensors, is described in details in Part A. Our
The pavement roughness levels can be classified according pavement roughness level algorithm is implemented based on
to the Technical Code of Maintenance for Urban Road CJJ36- the data from field experiments in Part B. Experimental results
2006 [20], which is one of the industry standards in the are discussed and presented in Part C.
People's Republic of China.
This industry standard shows that pavement roughness level A. Experimental setting
is evaluated by Riding Quality Index (RQI), which is a As mentioned above, RCM-TAGPS depends on the
comprehensive metric of driving comfort. The relationship accurate signals like acceleration and velocity. As a result, the
between RQI and IRI is described as follows: accelerometers and GPS sensors must be correctly placed so
that the acquired information is exact. Figure 5 shows the
RQI = 4.98-0.34×IRI (10) placement of these sensors in our experiments. The three-axis
The range of the value of RQIs varies from 0 to 5 normally. accelerometer is attached to the right side of dashboard, so that
We set RQI to 0 if the value calculated is negative. it can approach the vibrations of the driving vehicle as far as
possible. Furthermore, we need place the three-axis
Finally, we classify pavement roughness into four levels, accelerometer in a special direction, where its X-axis is with
including excellent, good, qualified and unqualified, with the the same direction as the driving direction, its Y-axis in the
evaluation standards listed in Table I. corresponding horizontal direction and the Z-axis in the
corresponding vertical direction. However, the GPS sensor is
TABLE I. EVALUATION STANDARDS FOR PAVEMENT ROUGHNESS LEVEL fixed to the left front windshield in the lateral part of the
vehicle, to obtain real-time signals in urban area.
Pavement
v(km/h) RQI
roughness level
RQI>3.6 excellent
3.0<RQI<3.6 good
v>80
2.5<RQI<3.0 qualified
0<RQI<2.5 unqualified
RQI>3.2 excellent
2.8<RQI<3.2 good
40<v<80
2.4<RQI<2.8 qualified
0<RQI<2.4 unqualified (a) three-axis accelerometer (b) GPS sensor
RQI>3.0 excellent Figure 5. Experimental setting of sensors
2.6<RQI<3.0 good
v<40
2.2<RQI<2.6 qualified To evaluate the pavement roughness level comprehensively
0<RQI<2.2 unqualified and effectively, we collect data from actual roads with various

1035
characters, including the following types: (I) smooth roads B. Evaluation
where the segments of road surface are considered smooth, (II) We clean the raw data with our data cleaning algorithm,
general roads where there are small particles such as sands or including outlier removal, data interpolation and so on. It is
stones, (III) roads with some potholes, (IV) roads with some noted that the average velocity of our experimental vehicles is
manholes, (V) roads with some decelerating belts. normally around 40 km/h because the heavy traffic in urban
We carried out hundreds of drives, and collected data by the area.
experimental vehicles equipped with RCM-TAGPS in urban Figure 6 shows the continuous three-axis accelerations
area of Beijing. The traces contain continuous acceleration and against the sampling points on various road types. Specially,
GPS data. Here, the acceleration consists of about 100 raw Figure 6(a) demonstrates three-axis acceleration on smooth
readings per second, while the GPS data is produced once per roads, while the others on general roads, roads with some
second, as described in Section III. potholes, roads with some manholes, and roads with some
decelerating belts separately.

(a) smooth roads (b) general roads (c) roads with some potholes

(d) roads with some manholes (e) roads with some decelerating belts

Figure 6. The continuous three-axis accelerations on various road types

TABLE II. PAVEMENT ROUGHNESS LEVELS OF VARIOUS ROAD TYPES

As discussed in Section III, we implement our pavement
roughness level algorithm on various road types respectively. Pavement rough-
road type IRI RQI
Firstly, PSD of pavement surface roughness can be presented ness level
as the Fourier transform of the auto-correlation function of smooth roads 4.90 3.32 excellent
three-axis acceleration. Then we analyze the roughness general roads 6.18 2.88 good
deviation and IRI. Finally, pavement roughness level is roads with potholes 12.33 0.79 unqualified
classified by Table I. Pavement roughness levels of various roads with manholes 8.24 2.18 unqualified
road types are listed in Table II. roads with decele-
6.81 2.66 qualified
rating belts

1036
C. Experiment results REFERENCES
Experimental results of RCM-TAGPS are discussed as [1] J. Eriksson, L. Girod, and B. Hull, etc. “The Pothole Patrol: Using a
follows. Mobile Sensor Network for Road Surface Monitoring”, 6th International
Conference on Mobile Systems, Applications, and Services (MobiSys
The Z-axis accelerometer should be more concerned. 2008), Breckenridge, USA, June, 2008, pp. 29-39.
During the experimental drives, the X-axis is with the same [2] R. Bishop, “A survey of intelligent vehicle applications worldwide”,
direction as driving, while the Y-axis and the Z-axis are in the IEEE Intelligent Vehicles Symposium (IV 2000), Dearborn, MI, USA,
May, 2000, pp. 25-30.
horizontal and vertical direction respectively. Obviously, the Z-
axis has a more prominent waveform on rough pavement, [3] X. Chu, and Y. Wu, “Desigend on the low cost system framwork of road
conditon recgonition based on roadside multi-sensoers”, Asia-Pacific
which should be investigated with priority. Confernce on Information Precessing (APCIP 2009), Guangdong, China,
July, 2009, pp. 173-176.
As illustrated in Figure 6(a), when driving on smooth roads,
[4] J. Budras, A synopsis on the current equipment used for measuring
the acceleration data are very smoothly with only small pavement smoothness. [Link]
fluctuations, so that the overall road condition is outstanding. [Link], August, 2001.
We infer from Figure 6(b) that there are some instantaneous [5] L. Sun, “Optimum design of road-friendly vehicle suspension systems
subjected to rough pavement surfaces”, Applied Mathmatical Modelling,
large pulses in the Z-axis acceleration on general roads. It 2002, vol. 26, pp. 635-652.
might be aroused by small particles or other interferences on
[6] L. Sun, and T. Kennedy, “Spectral analysis and parametic study of
the road, and the overall road condition is considered to be stochastic pavement loads”, Journal of Engineering Mechanics, 2002,
good. vol. 128, pp. 318-327.
[7] Y. Zhang, and R. Ma, “A study of pavement roughness measurement
Figure 6(c) shows the acceleration data sampled with system based on laser ranger finder”, International Conference on Image
continuous small potholes. The acceleration data are with great Analysis and Signal Processing (IASP 2009), Zhejiang, China, April,
fluctuations, and we deduce that the vehicles vibrated strongly 2009.
when passing by these potholes. The road condition is [8] J. Dyer, D. Dyer, and J. Devore, “Estimating International Roughness
unqualified in general. Index from noisy profilograph measurements”, International
Instrumentation and Measurement Technoloty Conference (I2MTC
Figure 6(d) indicates the acceleration data sampled with 2009), Singapore, May, 2009.
some manholes, which is much similar to Figure 6(c). The [9] R. Hostettler, W. Birk, and M. Nordenvaad, “Surface mounted vehicle
acceleration data are also with great fluctuations, and the road property sensing for cooperative vehicle infrastructure systems”, 16th
condition is unqualified. ITS World Congress and Exhibition on Intelligent Transport Systems
and Services, Stockholm, Sweden, September, 2009.
From Figure 6(e), we find there is a large pulse at certain [10] Y. Qing, A. Lim, and P. Agrawal, “GPSFR: GPS-free geographic
points in the Z-axis acceleration when passing by several routing protocol for intelligent vehicles”, International Conference on
narrow decelerating belts. The road condition is considered to Comunications and Networks in China (Chinacom2006), Beijing, China,
October, 2006, pp. 1-8.
be qualified in spite of these disturbances.
[11] A. Thiagarajan, L. Ravindranath, and K. Lacurts, etc. “VTrack: accurate,
energy-aware road traffic delay estimation using mobile phones”, 7th
VI. CONCLUSION ACM Conference on Embedded Networked Sensor Systems (SenSys
2009), Berkeley, California, USA, November, 2009, pp. 85-98.
This paper describes the design, implementation and [12] L. Sun, “Simulation of Pavement Roughness and IRI Based on Power
experimental evaluation of a road condition monitoring system, Spectral Density”, Mathematics and Computers in Simulation, 2003, vol.
which is called RCM-TAGPS. RCM-TAGPS uses mobile 61, pp. 77-88.
vehicles embedded with a three-axis accelerometer and a GPS [13] D. Xu, A. Mohamed, and R. Yong, etc. “Development of a criterion for
sensor to collect the data of three-axis acceleration and velocity road surface roughness based on power spectral density function”,
with a high frequency. We clean the sampling raw data with Journal of Terramechanics, 1992, vol. 29, pp. 477-486.
our clean algorithm, including outlier removal, data [14] W. Gardner, "A sampling theorem for nonstationoary random process",
interpolation and so on, to deal with technical challenges like IEEE Transactions on Information Theory, 1972, vol. 18, pp. 808-809.
unavailable GPS and transmission error. Then we analyze the [15] L. Cohen, “Generalization of the Wiener-Khinchin theorem”, IEEE
Signal Processing Letters, 1998, vol. 5, pp. 292-294.
Power Spectral Density (PSD) of pavement surface roughness,
[16] A. Fung, “A note on the Wiener-Khintchine theorem for
estimate the International Roughness Index (IRI), and classify autocorrelation”, Proceedings of the IEEE, 1967, vol. 55, pp. 594-595.
the pavement roughness levels into four levels according to the
[17] T. Semiha, and A. Huseyin, “A study of random vibration characteristics
Technical Code of Maintenance for Urban Road CJJ36-2006. of the quarter-car model”, Journal of Sound and Vibration, 2005, vol.
Finally we evaluate RCM-TAGPS with actual driving in urban 282, pp. 111-124.
area of Beijing. Experimental results show that RCM-TAGPS [18] C. Dodds, and J. Robson, "The description of road surface roughness",
can evaluate the pavement roughness level correctly, even with Journal of Sound and Vibration, 1973, vol. 31, pp. 175-183.
some interference like potholes, manholes and decelerating [19] S. Lou, Y. Wang, and C. Xu, “Study of calculation of IRI based on
belts. Furthermore, RCM-TAGPS consumes no more than 50 power spectral density of pavement surface roughness”, Journal of
dollars in total, which is only 1/4400 to 1/160 of these existing Highway and Transportantion Research and Development, 2007, vol. 24,
pp. 12-15.
systems [4] with a cost of about 8,000 to 222,000 dollars.
[20] Y. Shucong, Technical Code of Maintenance for Urban Road CJJ36-
Therefore, RCM-TAGPS is more likely to be widely used in 2006, Beijing, China Architecture & Building Press, 2006.
municipal engineering.

1037