Scientia Iranica B (2011) 18 (4), 930–937

Sharif University of Technology

Scientia Iranica
Transactions B: Mechanical Engineering
www.sciencedirect.com

Traffic condition recognition using the k-means clustering method

M. Montazeri-Gh, A. Fotouhi ∗
Systems Simulation and Control Laboratory, School of Mechanical Engineering, Iran University of Science and Technology, Tehran,
P.O. Box 16846-13114, Iran

Received 24 November 2010; revised 14 March 2011; accepted 7 June 2011

KEYWORDS Abstract This paper presents a methodological approach to traffic condition recognition, based on
Traffic condition driving segment clustering. Traffic condition recognition has many applications to various areas, such as
recognition; intelligent transportation, adaptive cruise control, pollutant emissions dispersion, safety, and intelligent
Hybrid electric vehicle; control strategies in hybrid electric vehicles. This study focuses on the application of driving condition
Driving feature; recognition to the intelligent control of hybrid electric vehicles. For this purpose, driving features
k-means clustering; are identified and used for driving segment clustering, using the k-means clustering algorithm. Many
Driving data collection; combinations of driving features and different numbers of clusters are evaluated, in order to achieve the
GPS. best traffic condition recognition results. The results demonstrate that traffic conditions can be correctly
recognized in 87 percent of situations using the proposed approach.
© 2011 Sharif University of Technology. Production and hosting by Elsevier B.V.
Open access under CC BY-NC-ND license.

1. Introduction Beside the above mentioned applications, traffic information

might be used in an intelligent Hybrid Electric Vehicle (HEV)
Traffic flow in each city or region contains different driving power management system. This application has also recently
conditions that take place repeatedly. The driving condition attracted research interest [13–19]. This study also aims at
varies from place to place or from public transportation, such driving conditions analysis for use in an adaptive HEV control
as buses, to private vehicles. Driving conditions are identified strategy, based on traffic condition recognition.
by their features, which have been investigated in previous A hybrid vehicle is a type of vehicle in which at least
studies [1,2]. two different power sources are used. A HEV includes an
Traffic flow modeling and driving condition analysis have
internal combustion engine with an associated fuel tank, and
many applications to various areas, such as Intelligent Trans-
an electric motor with its associated battery. In order to utilize
portation Systems (ITS), adaptive cruise control, pollutant emis-
the two power sources in a HEV effectively, a control strategy is
sions dispersion and safety. A group of studies in this area are
essential.
related to the mathematical modeling of traffic flow [3–6]. In
other studies, driving pattern recognition is utilized in order In an advanced type of HEV control strategy, the controller
to improve driving safety [7,8] and in some research, pollutant adapts itself to the current traffic condition to reduce Fuel
emissions dispersion in the air has been investigated [9,10]. In Consumption (FC) and exhaust emissions. Traffic Condition
addition, driving cycle development has been studied in previ- Recognition (TCR) is a critical sub-system of this intelligent
ous studies, based on driving condition analysis [11,12]. HEV control strategy. The TCR unit recognizes the current
traffic condition by classifying the driving segments, based on
a driving database. The HEV controller parameters are then
∗ Corresponding author. adapted, regarding the TCR, in order to reduce vehicle fuel
E-mail address: [email protected] (A. Fotouhi).
consumption and pollutant emissions. Details of the procedure,
1026-3098 © 2011 Sharif University of Technology. Production and hosting by using traffic condition information in the HEV control unit, is
Elsevier B.V. Open access under CC BY-NC-ND license. out of the framework of this paper. More descriptions of this
Peer review under responsibility of Sharif University of Technology. procedure can be found in some Refs. [13–19].
doi:10.1016/j.scient.2011.07.004 Although driving data is used in HEV control in some pre-
vious studies, no mathematical approach has been conducted
for driving data clustering. In much of the research, some driv-
ing cycles are used as representative of different driving con-
ditions [13,14,19]. The number of driving cycles is limited and
does not necessarily contain all driving conditions. In addition,
 M. Montazeri-Gh, A. Fotouhi / Scientia Iranica, Transactions B: Mechanical Engineering 18 (2011) 930–937 931

Figure 1: Schematic of a traffic condition recognition (TCR) system.

Figure 2: The AVL device.
each driving cycle belongs to a specific city or region and cannot
be used in other places. 16 h. After that, the data is transferred to a computer in which
In this study, driving data is investigated more deeply. the software of the device has been installed.
At first, driving data collection has been performed using In order to analyze the driving data, a partitioning approach
Automatic Vehicle Location (AVL) devices. Consequently, a large is proposed in this study, based on definition of the driving
number of driving segments are used in the investigations. segment. A ‘‘driving segment’’ is defined as part of the vehicle
Then, features of driving segments are extracted and analyzed velocity profile during a period of time. This partitioning
in order to identify the most proper driving features. The approach is helpful for extraction of driving features from
selected driving features are utilized for driving condition driving data. The length of the driving segment can be set to
recognition using a new approach based on driving segment any desired value. Here, regarding the application of the study
clustering. The k-means method is used for driving segment to HEV control, the length of the segment is set to 150 s. This
clustering. can be justified, regarding the nature of traffic data, which
A schematic representation of the developed Traffic Condi- varies very quickly. For example, traffic data for the past 30 min
tion Recognition (TCR) sub-system in this study is illustrated in may not be useful now. On the other hand, in very short time
Figure 1. As shown in the figure,TCR works based on classifica- intervals of less than one minute, the traffic condition cannot
tion of driving segments. In a TCR unit, vehicle speed is stored be recognized correctly because of instantaneous accelerating
and analyzed on-line in order to recognize the current traffic and decelerating. An approximately 3-min length for driving
condition. In the classification procedure, features of the last segments is also considered in other studies [13]. Figure 4
driving segment are extracted and the cluster of the segment is depicts a sample of measured driving data containing four
identified. The off-line segment consists of driving data gather- 150 (s) driving segments.
ing, driving features selection and driving segments clustering. In this study, 11 driving features are defined to characterize
The output of the off-line segment is a database used for TCR. the driving segments. The driving features are presented in
The structure of the paper is as follows. Section 2 explains Table 2, many of which are used in other studies [20]. It should
driving data gathering, segmentation and driving features. In be noted that the approach of this study is applicable to any type
Section 3, driving features are evaluated based on a perfor- and number of driving features.
mance index. The k-means clustering method is described in Regarding application of the study to the TCR sub-system of
Section 4. Application of the k-means method for the cluster- the HEV control unit, the use of all driving features for traffic
ing of driving segments is presented in Section 5. Finally, in condition recognition is time consuming, and is not applicable
Section 6, driving segments are clustered into a different num- under on-line conditions. So, in this study, an approach is
ber of groups, using various combinations of driving features, conducted to select more proper driving features as follows.
and results are analyzed.
3. Driving features evaluation
2. Driving data gathering, segmentation and driving fea-
tures In this section, the influence of driving features on vehicle
fuel consumption and exhaust emissions is investigated, using
In this study, Automatic Vehicle Location (AVL) devices computer simulations, in order to select the most effective
are employed for data collection. The AVL (Figure 2) is an features. Advanced Vehicle Simulator (ADVISOR) software [21]
advanced device for vehicle tracking and monitoring, which is utilized for vehicle simulation. The simulations are also
works based on Global Positioning System (GPS) technology. verified by experimental testing [22]. A common vehicle in Iran,
The driving data collection is performed by a light-weight called ‘‘Samand’’, is simulated on driving segments in order to
car called ‘‘Samand’’, moving in different parts of Tehran city calculate fuel consumption and emissions for each segment.
(Figure 3), during a six month period in 2009–2010 under Then, the relationship between driving features and vehicle
real traffic conditions. The driving data includes date/time, fuel consumption and emissions are evaluated using correlation
number of satellites, longitude, latitude, speed and altitude of coefficients [23] between them. Besides Fuel Consumption (FC),
the vehicle during every second. A sample of driving data is three exhaust emissions are considered, including unburned
presented in Table 1. The AVL device communicates with the hydrocarbons (HC), Carbon Monoxide (CO) and oxides of
GPS satellite by an antenna. The antenna sticks on the roof of nitrogen (NOx).
the vehicle and is connected to the device by a cable. The power A performance index is defined in order to bond all
of the AVL device is supplied by the car battery. The device turns correlation coefficients in one variable for each driving feature.
on automatically when the car switch is on and it begins to save In this way, the driving features can be compared to each
the data. The device memory is enough for data collection over other and consequently can be ranked based on performance
932 M. Montazeri-Gh, A. Fotouhi / Scientia Iranica, Transactions B: Mechanical Engineering 18 (2011) 930–937

Figure 3: Path in which driving data is collected in Tehran city.

Table 1: A sample of driving data collected by the AVL device.

UnitID Date/time Longitude Latitude Speed Heading Altitude Satellites ReportID Inputs Outputs

2009/10/11
1010000137 51.447183 35.751267 63 72 1364 4 1 0 0
14:35

2009/10/11
1010000137 51.447375 35.751317 66 71 1363 4 1 0 0
14:35

2009/10/11
1010000137 51.44757 35.751372 67 70 1363 4 1 0 0
14:35

2009/10/11
1010000137 51.447768 35.75143 69 69 1362 4 1 0 0
14:35

2009/10/11
1010000137 51.44797 35.751492 71 68 1362 4 1 0 0
14:35

2009/10/11
1010000137 51.448175 35.75156 73 67 1361 4 1 0 0
14:35

2009/10/11
1010000137 51.448383 35.751633 74 66 1361 4 1 0 0
14:35

2009/10/11
1010000137 51.448602 35.75171 78 66 1360 4 1 0 0
14:35

2009/10/11
1010000137 51.448817 35.751782 76 67 1360 4 1 0 0
14:35
··· ··· ··· ··· ··· ··· ··· ··· ··· ··· ···

index values. Some weights are also considered in the following and (COcc )j is the correlation coefficient between the jth driving
formula in order to enhance the importance of FC and/or one feature and CO, and so on. In addition, WFC , WCO , WNOx and wHC
of the emissions: Eq. (1) is given in Box I, where (FCcc )j is the are weights of FC, CO, NOx and HC, respectively. Figure 5 depicts
correlation coefficient between the jth driving feature and FC, correlation coefficients between vehicle fuel consumption
 M. Montazeri-Gh, A. Fotouhi / Scientia Iranica, Transactions B: Mechanical Engineering 18 (2011) 930–937 933

Table 2: Driving features.

Driving feature description Formula

i=1 vi where n is the length of the driving segment in sec, vi

1
∑n
1 Mean velocity (Vmean ): Average of speed during a driving segment Vmean = n
is the velocity value in second i
Variance of velocity (σ ): Variance of velocity during a driving segment σ = 1n ni=1 (vi − Vmean )2
∑
2
3 Maximum acceleration: Maximum value of acceleration during a driving –
segment
4 Minimum acceleration: Minimum value of acceleration during a driving –
segment
tidle
5 Idle time percentage: Percentage of time in which the vehicle stops ttotal
× 100 where tidle is the idle time and ttotal is the total time of driving
during a driving segment and speed is zero segment
6 Maximum velocity: Maximum value of speed during a driving segment –
7 Minimum velocity: Minimum value of speed during a driving segment –
8 (max velocity) − (min velocity): Maximum value minus minimum value –
of speed during a driving segment
tcruise
9 Cruise percentage: Percentage of time during a driving segment in which ttotal
× 100 where tcruise is the cruise time and ttotal is the total time of
the vehicle moves with a constant speed driving segment
ai for ai < 0 nd is the number of points that acceleration
1
∑nd
10 Average Decelerating (A.D.): Average value of negative accelerations A.D. = nd i=1
during a driving segment is negative
ai for ai > 0 na is the number of points that acceleration
1
∑na
11 Average Accelerating (A.A.): Average value of positive accelerations A.A. = na i=1
during a driving segment is positive

Figure 4: 150 (s) driving segments.

and 11 driving features. The correlation coefficients between

exhaust emissions and driving features are also presented in Figure 5: Correlation coefficients between fuel consumption and driving
Figure 6. The magnitude of a correlation coefficient indicates features.

the dependency between vehicle FC and/or emissions and that

driving feature. In other words, the greater the correlation
coefficient, the stronger the relationship is that exists between
the driving feature and the vehicle FC and/or emissions.
It is observed in Figure 5 that features, 1, 6 and 11
(i.e. mean velocity, maximum velocity and average accelerating,
respectively) have greater FC correlation coefficients, which
indicate a stronger relationship between these driving features
and FC. It comes from the results where FC depends on
these features more than other driving features. After the
3 above-mentioned driving features, the largest correlation
coefficient belongs to the 9th driving feature (i.e. cruise motion
percentage), and so on.
Concerning the relationship between exhaust emissions and
driving features, Figure 6 shows that the two emissions, CO
and HC, depend on the 1st and 6th driving features (i.e. mean
velocity and maximum velocity, respectively) more than the
others. However, the pollutant factor, NOx, depends mostly on
the 11th and 9th driving features (i.e. average accelerating and Figure 6: Correlation coefficients between exhaust emissions and driving
cruise motion percentage, respectively). features.
934 M. Montazeri-Gh, A. Fotouhi / Scientia Iranica, Transactions B: Mechanical Engineering 18 (2011) 930–937

WFC (FCcc )j + WCO (COcc )j + WNOx (NOxcc )j + WHC (HCcc )j

(Performance index)j = , for j = 1, . . . , 11, (1)
WFC + WCO + WNOx + WHC

Box I:

Table 3: Ranks of driving features based on performance index values and Step 3: Compute new cluster centers, z1∗ , z2∗ , . . . , zK∗ , as follows:
WFC = WCO = WNOx = WHC = 1.
1 −
Feature Feature name Performance index zi∗ = ∈ Ci xj , i = 1, 2, . . . , K , (4)
number WFC = WCO = WNOx = ni xj
WHC = 1

1 Mean velocity 0.40

where ni is the number of elements belonging to cluster
6 Maximum velocity 0.38 Ci .
11 Average accelerating 0.33 Step 4: If zi∗ = zi , i = 1, 2, . . . , K , then terminate. Otherwise,
5 Idle time percentage 0.31 continue from Step 2.
9 Cruise percentage 0.30
8 (max vel.) − (min vel.) 0.30 Note that, in cases where the process does not terminate at
3 Maximum acceleration 0.29 Step 4 normally, then, it is executed for the maximum number
7 Minimum velocity 0.19 of iterations.
2 Variance of velocity 0.18
10 Average decelerating 0.16
4 Minimum acceleration 0.11 5. Driving segments clustering

In the previous part, the k-means clustering method is

explained. In this section, the driving segments will be clustered
The driving features are sorted in Table 3 based on per-
using the k-means method. As presented in Step 1 of the
formance index values, using weights: WFC = WCO = k-means clustering method, at first the number of clusters
WNOx = WHC = 1. The results demonstrate that ‘‘mean should be determined. In this study, different numbers of
velocity’’, ‘‘maximum velocity’’, ‘‘average accelerating’’, ‘‘idle clusters are investigated. The points which should be clustered
time percentage’’, ‘‘cruise percentage’’ and ‘‘maximum ve- are driving segments, so parameter ‘‘n’’ is the number of driving
locity − minimum velocity’’ driving features have the most segments. For example, point x1 stands for the first segment.
influence on vehicle fuel consumption and emissions. In the fol- The dimension of each point depends on the number of features
lowing, these two driving features are used for driving segments used for segment clustering. If two features are used, the
clustering. Before the clustering of driving segments, a brief de- dimension of the clustering is two. If three features are used,
scription about the clustering method is presented in the next the dimension of the clustering is three, etc.
section. For selection of the driving features, the results of Section 3
are utilized, which demonstrate that ‘‘mean velocity’’, ‘‘maxi-
4. k-means clustering method mum velocity’’, ‘‘average accelerating’’, ‘‘idle time percentage’’,
‘‘cruise percentage’’ and ‘‘maximum velocity − minimum ve-
locity’’ are the most effective driving features, regarding their
In this section, the k-means clustering method is explained.
influence on vehicle fuel consumption and exhaust emissions.
Clustering in N-dimensional Euclidean space, RN , is the process
In this section we focus on the clustering of driving segments
of partitioning a given set of n points into K groups (or clusters)
using these above-mentioned driving features.
based on a similarity/dissimilarity metric [24]. Let the set of
Following feature evaluation and selection, driving segment
n points {x1 , x2 , . . . , xn } be represented by the set S, and K
clustering is performed to provide the driving condition groups
clusters be represented by C1 , C2 , . . . , CK . Then: used in TCR. Using two driving features for clustering driving
Ci ̸= ∅ for i = 1, . . . , K , segments, each driving segment can be plotted as a point in
2-dimensional feature space. For example, the scatter plot of
Ci ∩ Cj = ∅ for i = 1, . . . , K , j = 1, . . . , K , i ̸= j, driving segments in a feature space is presented in Figure 7,
and: using ‘‘mean velocity’’ and ‘‘average accelerating’’ driving
features. As a sample, driving segments are clustered into three
K
 clusters in this figure.
Ci = S . (2) All possible cases of driving segments clustering, due to
i=1 different selections of driving features and different numbers of
clusters, are investigated in this section. For this purpose, only
One of the most widely used clustering techniques available
the six previously mentioned driving features (stand at the top
in the literature is the k-means algorithm [25]. The k-
of Table 3) are used for the clustering exercise.
means algorithm attempts to solve the clustering problem by
Two criteria are employed for the selection of driving
optimizing a given metric. The steps of the k-means algorithm
features and the number of clusters. The first is the percentage
are briefly described in the following [24]: of correct TCRs. Because of the target of this study, which
Step 1: Choose K initial cluster centers, z1 , z2 , . . . , zK , randomly is TCR, for intelligent HEV control, the recognition result is
from the n points {x1 , x2 , . . . , xn }. very important and has been investigated for each case. The
percentage of correct recognition is formulated as follows:
Step 2: Assign point xi , i = 1, 2, . . . , n, to cluster Cj , j ∈
{1, 2, . . . , K }, if: Correct TCR Percentage
xi − zj  < xi − zp  ,
   
p = 1, 2, . . . , K , j ̸= p. number of correctly recognized segments
(3) = , (5)
number of all segments
 M. Montazeri-Gh, A. Fotouhi / Scientia Iranica, Transactions B: Mechanical Engineering 18 (2011) 930–937 935

Table 4: Mean of silhouette values for different clusters using two features.

Number of clusters
Driving 3 4 5 6 7 8 9 10
feature
numbers

1, 6 0.58 0.55 0.54 0.53 0.52 0.52 0.51 0.50

1, 11 0.54 0.51 0.49 0.47 0.46 0.45 0.45 0.44
1, 5 0.60 0.54 0.59 0.56 0.55 0.55 0.57 0.56
1, 9 0.37 0.39 0.41 0.39 0.40 0.39 0.39 0.40
1, 8 0.47 0.44 0.40 0.40 0.39 0.38 0.39 0.39
6, 11 0.53 0.49 0.48 0.48 0.46 0.46 0.45 0.44
6, 5 0.58 0.54 0.57 0.56 0.55 0.56 0.56 0.54
6, 9 0.33 0.39 0.40 0.37 0.37 0.37 0.38 0.37
6, 8 0.47 0.43 0.41 0.41 0.41 0.41 0.40 0.40
11, 5 0.88 0.50 0.49 0.51 0.45 0.58 0.56 0.19
11, 9 0.61 0.61 0.70 0.69 0.73 0.76 0.66 0.69
11, 8 0.45 0.42 0.41 0.41 0.40 0.38 0.38 0.38
Figure 7: Scatter plot of driving segments in (mean of velocity) − (average 5, 9 0.61 0.59 0.63 0.65 0.77 0.55 0.5 0.52
accelerating) feature space after clustering. 5, 8 0.55 0.60 0.55 0.55 0.54 0.49 0.32 0.31
9, 8 0.42 0.45 0.46 0.53 0.54 0.54 0.58 0.56

where a correctly recognized segment is a 10-s driving segment,

with the same cluster number as its previous 150-s segment.
In other words, the cluster number of each 10-s segment is Table 5: Percentage of correct traffic condition recognition for different
estimated by the cluster number of the last 150-s segment and, clusters using two features.
when the two cluster numbers are the same, recognition is Number of clusters
correct. It should be noted that each 10-s or 150-s segment is Driving 3 4 5 6 7 8 9 10
mapped to the driving feature space, in order to identify its feature
cluster number. numbers
The second criterion is the mean of the silhouette values 1, 6 80 72 67 61 57 53 49 46
of all points. The silhouette value for each point is a measure 1, 11 80 71 65 59 55 51 48 44
of how that point is similar to the points in its own 1, 5 81 72 68 61 56 52 50 47
1, 9 70 58 55 51 45 40 39 35
cluster, compared to the points in another cluster [26,27]. The 1, 8 79 70 63 59 55 52 49 47
silhouette value for an individual data item, ‘‘p’’, is defined as 6, 11 81 73 68 62 58 55 52 49
follows: 6, 5 82 75 70 65 60 58 57 51
6, 9 77 59 56 48 46 41 39 35
b(p) − a(p)
S (p) = p = 1 to number of all points, (6) 6, 8 79 73 65 61 57 55 51 49
max{a(p), b(p)} 11, 5 87 73 71 46 34 34 32 26
11, 9 47 37 33 32 30 30 23 22
where a(p) denotes the average distance from the pth point 11, 8 50 46 38 35 30 26 24 23
to other points in its cluster, and b(p) denotes the average 5, 9 63 45 36 32 29 25 23 22
distance from the pth point to the points in the closest other 5, 8 64 61 47 44 41 28 25 25
9, 8 47 37 34 30 28 26 23 23
cluster, which is defined as the one yielding minimal b(p). The
silhouette value is within the range of −1 to +1 and is to be
maximized. The mean of the silhouette values of points in all
clusters is used as a criterion for representing the goodness of results is also investigated. For this purpose, the percentage
the clustering. The larger the mean silhouette value, the better of correct recognitions is presented in Table 5. According
is the clustering performed. to Table 5, increasing the number of clusters leads to a
decrease in TCR precision. The features pair, 5 and 11 (i.e. idle
6. Results analysis time percentage and average accelerating) leads to the best
recognition result. Under the best condition in which three
In this section, different combinations of driving features are clusters are applied, using the 5th and 11th driving features, the
investigated. In each case, the silhouette value and correct TCR percentage of correct recognition is 87%.
percentage are computed. Consequently, some driving features TCR results in Table 5 demonstrate that driving feature pairs,
are suggested to be used in an intelligent HEV control strategy. (1, 6), (1, 11), (1, 5), (6, 11), (6, 5) and (11, 5) are more proper
In the first step, only two driving features are used. than the others. In other words, driving features 1, 5, 6, 11
According to the first row of Table 4, using ‘‘mean velocity’’ (i.e. mean velocity, maximum velocity, idle time percentage and
and ‘‘maximum velocity’’ driving features (feature numbers 1, average accelerating) are suitable for TCR. In the following, the
6), the largest silhouette value is obtained for three clusters. effect of adding more driving features for TCR is investigated.
So, it is better to cluster driving segments into three clusters Tables 6 and 7 present the silhouette value and percentage
when these two features are used. Similarly, in the 2nd and of correct TCR, respectively, using more than two features.
3rd rows of the table, three clusters are suggested based on The results demonstrate that adding more features does not
silhouette values. In the 4th row, using ‘‘mean velocity’’ and improve clustering or TCR results.
‘‘cruise percentage’’ driving features (feature numbers 1, 9), the Because of the on-line application of the TCR unit to HEV
largest silhouette value is obtained for five clusters. Similarly, in control and the importance of computation time, it is not
each row, the largest silhouette value is typesetted in boldface. suggested to use more than two driving features for TCR. So,
As mentioned before, this study aims at traffic condition the use of two driving features ‘‘idle time percentage’’ and
recognition. Therefore, the influence of clustering on TCR ‘‘average accelerating’’ (feature numbers 5 and 11) is suggested.
936 M. Montazeri-Gh, A. Fotouhi / Scientia Iranica, Transactions B: Mechanical Engineering 18 (2011) 930–937

Table 6: Mean of silhouette value for different clusters.

Number of clusters
Driving 3 4 5 6 7 8 9 10
feature
numbers

1, 5, 6 0.59 0.55 0.52 0.55 0.54 0.53 0.51 0.51

1, 5, 11 0.57 0.51 0.55 0.51 0.50 0.48 0.47 0.49
1, 6, 11 0.56 0.52 0.49 0.49 0.48 0.46 0.45 0.44
5, 6, 11 0.55 0.50 0.53 0.50 0.48 0.50 0.48 0.47
1, 5, 6, 11 0.58 0.53 0.50 0.52 0.51 0.49 0.47 0.47

100
cluster 1
90 cluster 2
cluster 3
80

70
Idle time percentage

60

50

40

30

20

10

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4
Average accelerating (m/s2)

Figure 8: Scatter plot of driving segments in idle time percentage—average accelerating feature space.

Table 7: Percentage of correct traffic condition recognition for different features, ‘‘idle time percentage’’ and ‘‘average accelerating’’, are
clusters. the most suitable driving features for TCR. Using these two fea-
Number of clusters tures the percentage of correct recognition is obtained as 87%.
Driving 3 4 5 6 7 8 9 10 For further work, in order to use the TCR sub-system in a HEV
feature control unit, the controller parameters should be optimized for
numbers
different clusters. Consequently, the HEV uses each controlling
1, 5, 6 81 74 67 63 58 55 50 49 mode in its corresponding traffic condition.
1, 5, 11 80 72 68 61 57 52 49 47
1, 6, 11 81 73 66 61 57 53 49 46
5, 6, 11 82 75 70 64 59 58 54 51 References
1, 5, 6, 11 81 74 67 63 58 54 50 48

[1] Ericsson, E. ‘‘Variability in urban driving patterns’’, Int. J. Transp. Res. Part
D, 5, pp. 337–354 (2000).
The clustering of driving segments into three clusters is [2] Ericsson, E. ‘‘Independent driving pattern factors and their influence
on fuel-use and exhaust emission factors’’, Int. J. Transp. Res. Part D, 6,
depicted in Figure 8. As mentioned before, in this case, the pp. 325–341 (2001).
percentage of correct recognition is obtained as 87%. [3] Mallikarjuna, Ch. and Ramachandra Rao, K. ‘‘Cellular automata model for
As mentioned before, this study focuses on the application heterogeneous traffic’’, Int. J. Adv. Transport., 43(3), pp. 321–345 (2009).
of driving and traffic condition information in an intelligent [4] Qiu, Y. and Chen, S. ‘‘A new nonlocal traffic flow model with anticipation-
dependent relaxation time, equilibrium velocity and variance’’, Int. J. Stat.
HEV control strategy. In this paper, the first parts of the Mech. Theory Exp., p. P12012 (2010).
study, including driving features analysis and driving segments [5] Kanai, M. ‘‘Two-lane traffic-flow model with an exact steady-state
clustering, are presented. The results of this paper might be solution’’, Int. J. Phys. Rev. E, 82(6), p. 066107 (2010).
used for designing an intelligent multi-mode HEV control [6] Kerner, B.S., Klenov, S.L. and Hiller, A. ‘‘Empirical test of a microscopic
three-phase traffic theory’’, Internat. J. Non-Linear Mech., 49, pp. 525–553
strategy based on traffic condition recognition. (2007).
[7] Langford, J. and Koppel, S. ‘‘Epidemiology of older driver crashes—
7. Conclusion identifying older driver risk factors and exposure patterns’’, Int. J. Transp.
Res. Part F, 9, pp. 309–321 (2006).
[8] Lee, J.D., Li, J.D., Liu, L.Ch. and Chen, Ch.M. ‘‘A novel driving pattern
In this study, a new approach for Traffic Condition Recog- recognition and status monitoring system’’, Pacific-Rim Symposium on
nition (TCR) is presented based on the clustering of driving Image and Video Technology, PSIVT, Springer Proceedings, 4319, Hsinchu,
segments. The k-means clustering method is utilized for the Taiwan, pp. 504–512 (2006).
[9] Dirks, K.N., Johns, M.D., Hay, J.E. and Sturman, A.P. ‘‘A semi-empirical
clustering of segments in driving feature space. Several com- model for predicting the effect of changes in traffic flow patterns on carbon
binations of driving feature are investigated in order to achieve monoxide concentrations’’, Int. J. Atmos. Environ., 37, pp. 2719–2724
the best results. The results demonstrate that the two driving (2003).
 M. Montazeri-Gh, A. Fotouhi / Scientia Iranica, Transactions B: Mechanical Engineering 18 (2011) 930–937 937

[10] Chan, T.L. and Ning, Z. ‘‘On-road remote sensing of diesel vehicle emissions [21] Markel, T. and Brooker, A. ‘‘ADVISOR: a systems analysis tool for advanced
measurement and emission factors estimation in Hong Kong’’, Int. J. Atmos. vehicle modeling’’, Int. J. Power Sources, 110, pp. 255–266 (2002).
Environ., 39, pp. 6843–6856 (2005). [22] Montazeri-Gh, M., Fotouhi, A. and Naderpour, A. ‘‘Driving features and
[11] Montazeri-Gh, M. and Naghizadeh, M. ‘‘Development of the tehran car their influences on hybrid electric vehicle’s fuel consumption and exhaust
driving cycle’’, Int. J. Environ. Pollut., 30(1), pp. 106–118 (2007). emissions’’, Int. J. Engine Res., 17, pp. 42–51 (2009).
[12] Kamble, S.H., Mathew, T.V. and Sharma, G.K. ‘‘Development of real-world [23] Cohen, J., Cohen, P., West, S.G. and Aiken, L.S., Applied Multiple Regres-
driving cycle: case study of pune, india’’, Int. J. Transp. Res. Part D, 14, sion/Correlation Analysis for the Behavioral Sciences, 3rd ed., Lawrence Erl-
pp. 132–140 (2009). baum Associates, Hillsdale, NJ (2003).
[13] Lin, C.C., Jeon, S., Peng, H. and Lee, J.M. ‘‘Driving pattern recognition for [24] Hartigan, J.A., Clustering Algorithms, Wiley, New York (1975).
control of hybrid electric trucks’’, Int. J. Veh. Syst. Dyn., 42(41-58) (2004). [25] Rousseeuw, P.J. ‘‘Silhouettes: a graphical aid to the interpretation and
[14] Langari, R. and Won, J.S. ‘‘Intelligent energy management agent for a validation of cluster analysis’’, Int. J. Comput. Appl. Math., 20, pp. 53–65
parallel hybrid vehicle-part I: system architecture and design of the (1987).
driving situation identification process’’, Int. J. IEEE Trans. Veh. Technol., [26] Kaufman, L. and Rousseeuw, P.J., Finding Groups in Data: An Introduction to
54(3), pp. 925–934 (2005). Cluster Analysis, Wiley (1990).
[15] Won, J.S. and Langari, R. ‘‘Intelligent energy management agent for a [27] Tou, J.T. and Gonzalez, R.C., Pattern Recognition Principles, Addison-Wesley,
parallel hybrid vehicle-part II: Torque distribution, charge sustenance Reading, MA (1974).
strategies, and performance results’’, Int. J. IEEE Trans. Veh. Technol., 54(3),
pp. 935–953 (2005).
[16] Musardo, C., Rizzoni, G. and Staccia, B. ‘‘A-ECMS: an adaptive algorithm
Morteza Montazeri-Gh established the Systems Simulation and Control
for hybrid electric vehicle energy management’’, 44th IEEE Conference on
Laboratory at Iran University of Science and Technology (IUST) in 1996, where
Decision and Control, Seville, Spain, December, pp. 12–15 (2005).
he is currently Associate Professor of Mechanical Engineering. One of his current
[17] Montazeri-Gh, M. and Asadi, M. ‘‘Influence of the road grade on the
research projects is directed toward the use of driving data in intelligent HEV
optimization of fuzzy base HEV control strategy’’, Int. J. Soc. Automot. Eng.,
control.
SAE, Technical Paper Series, No. 2006-01-3293 (2006).
His main research interests include: Systems Simulation and Control.
[18] Chen, B.Ch., Wu, Y.Y. and Hsieh, F.Ch. ‘‘Design of multi-mode switch
strategy for lean burn engine using driving pattern recognition technique’’,
IEEE Conference on Systems, Man and Cybernetics, 2, Taipei, Taiwan,
pp. 1022–1029 (2006). Abbas Fotouhi is a Ph.D. candidate at Iran University of Science and Technology
[19] Montazeri-Gh, M. and Asadi, M. ‘‘Intelligent approach for parallel HEV (IUST) in Tehran, Iran, where since 2006 he has been working as a research
control strategy based on driving cycles’’, Internat. J. Systems Sci., 42(2), assistant at the Systems Simulation and Control Laboratory in the Department
pp. 287–302 (2011). of Mechanical Engineering. He has B.Sc. degree in railway rolling stock
[20] Brundell-Freij, K. and Ericsson, E. ‘‘Influence of street characteristics, driver engineering from IUST, and M.Sc. degree in Mechanical Engineering from
category and car performance on urban driving patterns’’, Int. J. Transp. Res. Tehran University. His research interests include hybrid electric vehicles, traffic
Part D, 10, pp. 213–229 (2005). conditions and driving features.