540 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO.

2, APRIL 2014

Vehicle Reidentification With Self-Adaptive Time

Windows for Real-Time Travel Time Estimation
Jiankai Wang, Nakorn Indra-Payoong, Agachai Sumalee, and Sakda Panwai

Abstract—This paper proposes a vehicle reidentification (VRI) [4], [5]. To overcome this difficulty, considerable attention
system with self-adaptive time windows to estimate the mean has been paid to using the emerging sensing technologies to
travel time for each time period on the freeway under traffic directly track the individual probe vehicle and, hence, collect
demand and supply uncertainty. To capture the traffic dynamics
in real-time application, interperiod adjusting based on the expo-
the associated travel time (which could be termed as the probe-
nential smoothing technique is introduced to define an appropriate vehicle-based method). Various advanced technologies, such
time-window constraint for the VRI system. In addition, an in- as Bluetooth [6], Global Positioning System technologies [7],
traperiod adjusting technique is also employed to handle the non- license plate recognition technique [8], and cellular phones [9],
predictable traffic congestion. To further reduce the negative effect have been incorporated to assign a unique identity (e.g., plate
caused by the mismatches, a postprocessing technique, including number, media access control address, and radio frequency
thresholding and stratified sampling, is performed on the travel
time data derived from the VRI system. Several representative
identification tag) to the probe vehicle. By accurate match-
tests are carried out to evaluate the performance of the proposed ing of vehicle identities, the travel time of the probe vehicle
VRI against potential changes in traffic conditions, e.g., recurrent can be directly measured. Although these probe-vehicle-based
traffic congestion, freeway bottlenecks, and traffic incidents. The approaches appear promising for travel time estimation, their
results show that this method can perform well under traffic successes rely on a high penetration rate of probe vehicles. In
demand and supply uncertainty. addition, vehicle tracking based on the unique identity could
Index Terms—Stratified sampling, time window, travel time raise privacy concerns. In this case, the vehicle reidentification
estimation, vehicle reidentification (VRI), video image processing (VRI) scheme, which does not intrude the driver’s privacy,
(VIP) systems. provides an alternative way to measure travel time.
Generally, VRI is a process of matching vehicle signatures
I. I NTRODUCTION (e.g., waveform [10], vehicle length [11], [12], vehicle color
[13], [14], and partial plate number [15], [16]) from one de-
A S ONE of the best indicators for evaluation of the perfor-
mance of the traffic system, accurate travel time data are
crucial for efficient traffic management and transport planning.
tector to the next in the traffic network. On one hand, the
nonuniqueness of the vehicle signature would allow the VRI
In addition, the individual travelers also require such informa- system to track the vehicle anonymously. Moreover, the pene-
tion (e.g., mean travel time) to make a better route decision tration rate is 100% in principle as no in-vehicle equipment is
for their journeys. Therefore, it is of great importance to es- required. On the other hand, this property of very nonunique-
timate the mean travel time in a robust and accurate manner. ness imposes a great challenge on the development of the
Because of the worldwide deployment of inductive loops, a vehicle signature matching method. To improve the matching
large number of studies focused on utilizing the traffic data accuracy, Coifman [17] compared the lengths of vehicle pla-
(e.g., spot speed and traffic flow) obtained from the traditional toons (i.e., vehicle platoon matching method). To further con-
sensors to indirectly measure the mean travel time [1]–[3]. sider the noise and the nonuniqueness of the vehicle signature,
Despite their computational efficiency and analytical simplicity, Kwong et al. [18], [19] proposed a statistical matching method
these indirect methods based on traditional sensors would result in which the vehicle signature is treated as a random vari-
in large errors when it comes to serious traffic congestion able, and a probabilistic measure is introduced for matching
decision-making. The aforementioned approaches, however,
Manuscript received April 13, 2013; revised July 18, 2013; accepted
are limited to the case with only one lane arterial and have a set
September 2, 2013. Date of publication October 25, 2013; date of current of stringent assumptions on vehicle traveling behaviors (e.g.,
version March 28, 2014. This work was supported by the Hong Kong Research no overtaking and no lane changing). Sumalee et al. [20] ex-
Grants Council under General Research Fund PolyU 5242/12E. This work tended the statistical signature matching method illustrated by
is also part of a collaborative research with the Division of Research and
Development of Expressway Authority of Thailand. The Associate Editor for Kwong et al. [18] to a more practical case in which overtaking
this paper was W.-H. Lin. (Corresponding author: A. Sumalee.) between vehicles and vehicle matching across multiple lanes
J. Wang and A. Sumalee are with the Department of Civil and Environmental are both allowed. This proposed VRI system, which is also
Engineering, The Hong Kong Polytechnic University, Kowloon, Hong Kong
(e-mail: [email protected]; [email protected]). referred to as the basic VRI, is based on the emerging video
N. Indra-Payoong is with the Department of Logistics and Supply image processing (VIP) systems [21], which enjoy several
Chain Management, Burapha University, Bangkok 20131, Thailand (e-mail: advantages over other traffic detectors (e.g., inductive loops and
[email protected]).
S. Panwai is with the Expressway Authority of Thailand (EXAT), Bangkok magnetic sensors) [22] as follows.
10900, Thailand (e-mail: [email protected]). • First, VIP systems are capable of monitoring multiple
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. lanes and can function as zone detectors rather than point
Digital Object Identifier 10.1109/TITS.2013.2282163 sensors (e.g., magnetic sensor and inductive loops).
1524-9050 © 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
WANG et al.: VRI WITH SELF-ADAPTIVE TIME WINDOWS FOR REAL-TIME TRAVEL TIME ESTIMATION 541

• Unlike the inductive loops, VIP systems could provide us tential changes in traffic conditions. Interperiod adjusting
the speed-independent signature. Various detailed vehicle of the time window based on the exponential smooth-
features (e.g., vehicle color, length, and type), which are ing technique is adopted to capture the traffic dynamics
independent of the vehicle speed, can be extracted. from period to period, whereas intraperiod adjusting is
A probabilistic data fusion rule was then introduced to combine employed to handle the nonpredictable traffic congestion
these features derived from VIP systems to generate a match- (e.g., caused by traffic incidents or the bottleneck effect).
ing probability (posterior probability) for matching decision- After the theoretical development, various numerical tests are
making. Sumalee et al. [20] also introduced a prior (fixed) time conducted to demonstrate the application of the improved VRI
window, which sets the upper and lower bounds of the travel system. The first simulation test investigates the feasibility of
time in the hope of ruling out the unlikely candidate vehicles utilizing the improved VRI system to estimate the mean travel
and, hence, improving the matching accuracy, which, in turn, time for a closed freeway segment containing recurrent conges-
would yield a more reliable travel time estimator. However, it tion due to exceeding traffic demand. In the second simulation,
is noteworthy that this basic VRI was specifically designed for the method is evaluated on a freeway corridor with on- and
a short time period in which the traffic condition is relatively off-ramps. A freeway bottleneck then arises due to the high
stable (i.e., steady-state condition) and may not be applicable merging demand and lane drops. The simulation results show
for “real time” application. that the proposed method performs well under the bottleneck
The development of the VRI system for “real time” imple- effect. The third simulation test is then conducted to test the
mentation is still difficult as it faces the following two major performance of the algorithm under nonrecurrent congestion
challenges. First, due to the traffic demand and supply un- (caused by traffic incidents).
certainty (e.g., fluctuation in travel demand, bottleneck effect, The rest of this paper is organized as follows: Section II
and traffic incidents), the traffic condition may substantially presents a brief review on the basic VRI system. In Section III,
change from period to period (i.e., free flow to congested). the postprocessing technique (e.g., thresholding and stratified
Under these circumstances, the fixed time-window constraint sampling) regarding the individual travel time obtained from
may compromise the performance of the basic VRI system. the VRI system is introduced. The detailed description and
Thus, instead of explicitly incorporating the time window, analysis of the self-adaptive time-window component is pro-
Lin and Tong [23] utilized the travel time information estimated posed in the following section. In Section V, simulated tests
from the spot speed data and proposed a combined estimation are carried out to evaluate the performance of the proposed
model to reidentify the vehicles. As previously discussed that system. Finally, we close this paper with the conclusions and
the travel time estimated from spot speed data is not reliable, future works.
this approach may not perform well under demand and supply
uncertainty. Second, vehicle mismatches, which are caused by II. BASIC V ISION -BASED VRI S YSTEM
the nonuniqueness of the vehicle signature and the complex
The basic vision-based VRI with a fixed time window is
topological structure of the traffic network, are to be expected.
devised to estimate the mean travel time under static traffic
This situation could be worse when a traffic incident happens.
conditions (e.g., a steady state of free flow/congestion). In
Therefore, a robust postprocessing technique regarding the
line with the other traditional VRI schemes, the basic vision-
individual travel time obtained from the VRI system is required.
based VRI also involves two major steps: 1) vehicle feature
Ndoye et al. [24] suggested a data clustering method to filter out extraction; 2) vehicle signature matching method.
the erroneous individual travel time caused by the mismatches.
For practical implementation, however, it may still be difficult
to distinguish between the correct and erroneous travel time A. Vehicle Feature Extraction
under “abnormal” traffic conditions (e.g., the occurrence of an As the detailed traffic data (particularly the vehicle feature
incident or the bottleneck effect). data) are not readily obtainable from the raw video record,
To this end, the objective of this paper is to propose an various VIP techniques are then employed for extracting the
improved VRI system to cope with the real-time travel time required information.
estimation purpose. Specifically, this study aims to estimate the 1) Vehicle Detection: The success of vehicle detection
mean travel time for each time period (e.g., 5-min period) on largely depends on the degree that a moving object (vehicle) can
the freeway under traffic demand and supply uncertainty. The be distinguished from its surroundings (background). In light
proposed VRI system is based on authors’ previous work [20] of this, background estimation technology is employed in the
with two major improvements as follows. detection subsystem. By calculating the media of a sequence of
• First, to filter out the erroneous travel time caused by the video frames, the background of the video image is obtained.
mismatches, a thresholding process based on the matching Then, the image segmentation technique is performed to iden-
probability is performed. A stratified sampling technique tify the foreground object (vehicle). The still image including
based on the vehicle type is then introduced to reduce the the detected vehicle is then clipped and stored for further
bias in the mean travel time estimates. feature extraction. Along with the detection of the vehicle, the
• Second, a self-adaptive time-window component (i.e., in- associated arrival time t and spot speed v are also collected.
terperiod and intraperiod adjusting) is introduced into the The normalized height of the vehicle image is adopted for
basic VRI system to improve its robustness against po- representing vehicle length L.
542 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

2) Vehicle Color Recognition: Color is one of the most natures (XiU , XjD ), the Bhattacharyya distance [26] is utilized
essential features for characterizing a vehicle. To reduce the to calculate the degree of similarity between color features, i.e.,
negative effect of illumination changes, hue–saturation–value  1/2
360 
(HSV) color space is adopted to represent the vehicle image.
First, the general RGB color images are converted into HSV dcolor (i, j) = 1 − CiU (k) · CjD (k) (2)
k=1
color model-based images. Hue and saturation values are then
exploited for color detection, whereas V (value) information where k denotes the kth component of the color feature vector.
is separated out from the color space. Second, a 2-D color The L1 distance measure is introduced to represent the differ-
histogram C is formed to represent the distribution (frequency) ence between the type feature vectors, i.e.,
of colors across a vehicle image. To be more specific, the
hue and saturation channels are divided into 36 and 10 bins, 
6
 U 
dtype (i, j) = Si (k) − SiD (k) . (3)
respectively. Thus, a color feature vector C with 360 elements
k=1
is obtained.
3) Vehicle Type Recognition: The vehicle type feature pro- The length difference is given by
vides other important information to describe a vehicle. In  
this research, the template matching method [25] is utilized dlength (i, j) = LU D
i − Lj . (4)
to recognize vehicle type. This method uses the L2 distance
metric to measure the similarity between the vehicle image However, in practice, it is unnecessary to compute the distance
and template images. Specifically, vehicles are classified into between all pairs of upstream and downstream vehicle signa-
six categories. For each category, the corresponding template tures. To rule out the unlikely candidate vehicles at the upstream
image is built for each lane. Finally, the normalized similarity database and improve the overall computational efficiency, a
value between the vehicle image (I) and the kth template image time-window constraint is introduced.
(T ) is given by 1) Time-Window Constraint: A time window, which sets
the upper and lower bounds of travel time, is introduced to

M 
N define the search space (i.e., set of potential upstream matches)
|I(m, n) − T (m, n)|2 for the downstream vehicle. Given a downstream vehicle j ∈
m=1 n=1
S(k) = (1) {1, 2, . . . , M }, its search space, i.e., S(j), is given by
GMN

S(j) = i|tD j − tmax ≤ ti ≤ tj − tmin
U D
(5)
where G denotes the maximum gray level (255); M and N
are the dimensions of the vehicle image. Thus, the vehicle where tmax and tmin are, respectively, the upper and lower
type/shape feature S is a 6-D vector that consists of the sim- bounds of the time window. For a sequence of downstream
ilarity score for each template. vehicles {1, 2, . . . , M }, the set of the candidate upstream ve-
In other words, a vehicle signature, i.e., X = {C, S, L}, is hicles, i.e., S, is defined as
generated for each detected vehicle, where C and S are the
M
normalized feature vector and the type (shape) feature vector,
S= S(j). (6)
respectively; L denotes the vehicle length. As previously men-
j=1
tioned, the associated arrival time t and spot speed v are also
obtained during the detection process. Therefore, the individual Under the static traffic condition, the time window [tmin , tmax ]
vehicle record can then be represented as (t, v, X). can be calibrated from the available historical travel time data.
With the associated search space S, the vehicle signature
matching method is equivalent to finding the correspondence
B. Vehicle Signature Matching Method between {1, 2, . . . , M } and S. Herein we introduce an indicator
For practical implementation, the vehicle records detected at variable to represent the matching result, i.e.,
the upstream station will be stored in the upstream database. 1, downstream vehicle j matches
Let Ui = (tU U U
i , vi , Xi ) denote the record of the ith upstream xij = upstream vehicle i ∈ S (7)
vehicle, where Xi = {CiU , SiU , LU
U
i } represents the associated 0, otherwise.
vehicle signatures (i.e., color, type, and length). In this case, the
upstream database is denoted as U = {Ui |i = 1, 2, . . .}, which Recall that for each pair of vehicle signatures, (XiU , XjD ),
could be updated with time propagation. Let D = {Dj |Dj = i ∈ S, j ∈ {1, 2, . . . , M }, one may compute the distance
j , vj , Xj ), j = 1, 2, . . . , M } denote the M vehicle records
(tD D D
(dcolor (i, j), dtype (i, j), dlength (i, j)) based on (2)–(4). A sim-
generated at the downstream station during a specific time pe- ple solution (i.e., distance-based method) is then to find the
riod (e.g., 5-min period). The VRI is to find the corresponding matching result xij with the minimum feature distance. How-
upstream vehicles for these M downstream vehicles based on ever, it should be noted that the vehicle signatures contain
the generated vehicle signatures. potential noise and are not unique. Therefore, the distance mea-
In order to quantify the difference between each pair of sure cannot really reflect the similarities between the vehicles.
upstream and downstream vehicle signatures, several distance Instead of directly comparing the feature distances, this study
measures are then incorporated. Specifically, for a pair of sig- utilizes the statistical matching method. Based on the calculated
WANG et al.: VRI WITH SELF-ADAPTIVE TIME WINDOWS FOR REAL-TIME TRAVEL TIME ESTIMATION 543

feature distance (dcolor (i, j), dtype (i, j), dlength (i, j)), a match- data set that contains a number of pairs of correctly matched
ing probability P (xij = 1|dcolor , dtype , dlength ) is provided for vehicles are utilized for estimating pdfs p1 and p2 . Likewise,
the matching decision-making. the likelihood functions for the type and length distances can
2) Calculation of Matching Probability: The matching be also obtained in a similar manner.
probability, which is also referred to as the posterior probabil- 4) Data Fusion Rule: In this paper, the logarithmic opinion
ity, plays a fundamental role in the proposed VRI system. By pool (LOP) approach is employed to fuse the individual likeli-
applying the Bayesian rule, one may have hood functions. The LOP is evaluated as a weighted product of
the probabilities, and the equation is given by
P (xij = 1|dcolor , dtype , dlength )
p(dcolor , dtype , dlength |xij = 1)P (xij = 1) p(dcolor , dtype , dlength |xij )
= (8)
p(dcolor , dtype , dlength ) 1
= p(dcolor |xij )α p(dtype |xij )β p(dlength |xij )γ ,
ZLOP
where p(dcolor , dtype , dlength |xij = 1) is the likelihood func-
tion; P (xij = 1) is the prior knowledge about the matching α+β+γ =1 (13)
result without observing the detailed vehicle feature data. In
where fusion weights α, β, and γ are used to indicate the degree
addition, one may also have
of contribution of each likelihood function, and ZLOP is the
p(dcolor , dtype , dlength ) normalizing constant. The weights can be also calibrated from
the training data set. By substituting (9), (10), and (13) into (8),
= p(dcolor , dtype , dlength |xij = 1)P (xij = 1) the desired matching probability can be obtained. For the sake
+ p(dcolor , dtype , dlength |xij = 0)P (xij = 0). (9) of simplicity, let Pij denote the matching probability between
upstream vehicle i ∈ S and downstream vehicle j.
Based on (8) and (9), it is observed that the calculation of 5) Bipartite Matching Method: Recall that the basic VRI
the matching probability is dependent on the deduction of the system is to find the matching result xij between the down-
likelihood function and the prior probability. In this particular stream vehicle set {1, 2, . . . , M } and its search space S
case, the prior probability is approximated by the historical (assume that there are N candidate vehicles) simultane-
travel time distribution, i.e., ously based on matching probability {Pij |i = 1, 2, . . . , N ; j =
1, 2, . . . , M }. The signature matching problem is then formu-
f (t(i, j))
P (xij = 1) = × 0.5 (10) lated as
η
f (t(i, j)) 
N 
M
P (xij = 0) = 1 − × 0.5 (11) min −Pij xij (14)
η x
i=1 j=1

where f (.) denotes the historical travel time distribution, t(i, j) s.t. xij ∈ {0, 1}, ∀i ∈ S, j ∈ {1, 2, . . . , M } (15)
is the time difference between upstream vehicle i and down-
stream vehicle j, and η is the normalizing factor. 
N
xij = 1, ∀j ∈ {1, 2, . . . , M } (16)
The calculation of the likelihood function is completed in i=1
two steps. First, individual statistical models for the three
feature distances are constructed, and the corresponding like- 
M
xij ≤ 1, ∀i ∈ S. (17)
lihood functions are also obtained (i.e., p(dcolor |xij = 1),
j=1
p(dtype |xij = 1), and p(dlength |xij = 1)). Then, a data fusion
rule is employed to provide an overall likelihood function used Objective (14) is to maximize the overall matching probabil-
in posterior probability (8). ities between the two sets. Constraint (15) ensures that the
3) Statistical Modeling of Feature Distance: Without loss decision variables are binary integers. Constraint (16) requires
of generality, only the probabilistic modeling of color feature that a downstream vehicle can have one matched vehicle at
distance is described. In the framework of statistical modeling, the upstream station, whereas constraint (17) guarantees that
the distance measure is assumed to be a random variable. Thus, an upstream vehicle can have, at most, one matched vehicle
for a pair of color feature vectors (CiU , CjD ), the distance at downstream (normally N > M ). This combinatorial opti-
dcolor (i, j) follows a certain statistical distribution. The con- mization problem is equivalent to a minimum-weight bipartite
ditional probability (i.e., likelihood function) of dcolor (i, j) is matching problem, which has already been widely studied
then given by and can be efficiently solved by the successive shortest path
p1 (dcolor (i, j)) , if xij = 1 algorithm with computational complexity of O(M 2 N ).
p (dcolor (i, j)|xij ) = (12) 6) Discussion on the Application of the Basic VRI System:
p2 (dcolor (i, j)) , if xij = 0
The detailed implementation of the basic VRI for mean travel
where p1 denotes the probability density function (pdf) of time estimation (e.g., from 10:00 A . M . to 10:05 A . M .) is
distance dcolor (i, j) when color feature vectors CiU and CjD summarized in the following flowchart (see Fig. 1). First, the
belong to the same vehicle, whereas p2 is the pdf of the distance system will initialize time stamp t and check whether a vehicle
dcolor (i, j) between different vehicles. A historical training is detected at the upstream and/or downstream stations. The
544 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

Fig. 2. Travel time for different vehicle types.

Fig. 1. Illustrative example of the basic VRI system.
(i.e., postprocessing and self-adaptive time window) can
be incorporated to improve the overall performance.
generated upstream vehicle records are stored in the upstream
vehicle database. Once a vehicle is detected at the downstream
station, the candidate vehicle set will be selected based on the III. P OSTPROCESSING T ECHNIQUE
time-window constraint. Meanwhile, the matching probability Upon completion of the basic VRI system, the raw travel
for each pair of vehicles is calculated. When current time t time for the jth, j ∈ {1, 2, . . . , M } downstream vehicle during
reaches 10:05 A . M ., the bipartite matching process based on the the evaluation period can be obtained and denoted as trj . Thus,
matching probability will begin, and the travel time data can the mean travel time without postprocessing is given by
be obtained. A detailed implementation of this system can be
1  r
M
found in [20]. For the aforementioned framework, the following
four comments should be taken into account. μr = t . (18)
M j=1 j
• First, it is noteworthy that the calculation of the match-
As the mismatches due to the nonuniqueness of the vehicle sig-
ing probability can be simultaneously performed along
nature are inevitable, the raw travel time may include erroneous
with the vehicle detection process at the downstream
information. Hence, accordingly, the estimator μr may not be
site (i.e., during the 5-min time period). In addition, the
reliable in practice.
bipartite matching process can be carried out efficiently
as explained before. Thus, the basic VRI can be imple-
mented in real time (which will be explained in detail in A. Stratified Sampling Technique
Section V-A).
One natural method is to perform thresholding on the raw
• Second, it is observed that the basic VRI heavily depends
travel time data based on the associated matching probability
on the specification of the time window. When a large
in an attempt to rule out the mismatches. However, another
time window is applied, search space S would include too
problem (i.e., biased estimation) may arise along with this
many candidate vehicles, which could lead to a significant
thresholding process. It is commonly believed that the travel
increase in computational time. On the other hand, a rel-
time of vehicles of different types (e.g., small cars and long
atively smaller time window may enable the algorithm to
trucks) is significantly different. Fig. 2 shows the travel time of
find the corresponding vehicle more efficiently; however,
different vehicle types fitted by a normal distribution, where ve-
it may also wrongly exclude the correct match from search
hicle type 1 denotes smaller cars, and vehicle type 2 represents
space S.
long trucks. These ground truth travel time data are collected
• Third, by using the historical travel time distribution to
from a freeway segment in Bangkok, Thailand. The authors also
approximate the prior knowledge (P (xij = 1)), one may
conducted various hypothesis tests (i.e., t-tests) to validate this
obtain a more reliable matching probability. In other
assumption. Thus
words, the basic VRI accepts the predefined time window
and the historical travel time distribution as exogenous H0 : μtype1 = μtype2 vs. Hα : μtype1 = μtype2
inputs and then perform the vehicle matching method.
Both of these two inputs can be derived from the mean where μtype1 and μtype2 are the mean travel time of small
travel time data (which will be explained in Section IV). cars and long trucks, respectively. The results show that the
• Fourth, the basic VRI cannot work well under traffic de- null hypothesis should be rejected, which means that the travel
mand and supply uncertainty, as the time window and prior times of different types of vehicles are “statistically” different.
knowledge may not be well defined. From the perspective In view of this, to further reduce the bias in mean travel time
of mean travel time estimation, two novel components estimation, the stratified sampling technique [27] based on the
WANG et al.: VRI WITH SELF-ADAPTIVE TIME WINDOWS FOR REAL-TIME TRAVEL TIME ESTIMATION 545

vehicle type is proposed. Specifically, the raw travel time data If vehicle j and upstream vehicle i∗ are truly matches, then the
{trj |j = 1, 2, . . . , M } are divided into two strata (i.e., small-car (2)
ratio between Pi∗ j and Pj is expected to be relatively larger.
stratum and long-truck stratum). The thresholding processes are Based on this basic idea, a predefined threshold value τ > 1 is
independently performed on these two strata. The final mean then imposed on this distinctiveness value, i.e.,
travel time μ is then computed as the weighted average of the ⎧
mean travel time over all vehicle type strata. The equation is ⎨ Pi∗ j /Pj(2) > τ, travel time tjk is retained for
then given as stratified sampling (21)
⎩
⎛ ⎞ Otherwise, travel time tjk is discarded.
2
Mk ⎝ 1 nk
μ= tjk ⎠ (19) By applying rule (21), the erroneous individual travel time data
M nk j=1 are expected to be identified and ruled out.
k=1

where Mk is the number of vehicles of type k, nk denotes the

sample size of vehicles of type k after the thresholding process, IV. S ELF -A DAPTIVE T IME -W INDOW C ONSTRAINT
and tjk is the travel time of the jth vehicle of type k after the Although the basic VRI is improved to some extent by
thresholding process. The design of the thresholding process imposing the postprocessing technique on the raw travel time
becomes the major concern in the following section. data (see Section III), it still cannot perform well under traffic
demand and supply uncertainty. (Some preliminary results are
B. Thresholding Process presented in Section V.) As mentioned in Section II, the basic
VRI heavily depends on the specification of two exogenous
For each individual vehicle stratum, the thresholding process inputs, i.e., time window and prior knowledge. Therefore, to
is independently performed. As explained before, one of the out- further improve the robustness of the VRI system against
puts of basic VRI is the matching result xij , whereas the other potential changes in traffic conditions, these two inputs should
output is the associated matching probability Pij . The overall be adjusted accordingly (i.e., self-adaptive).
idea of thresholding is to apply a certain rule to these outputs Intuitively, the time window can be derived from the travel
(i.e., xij and Pij ) in order to identify the associated erroneous time data (i.e., travel time distribution). Given the mean value
travel time. μt and the variance σt2 of the travel time distribution during
For downstream vehicle j in vehicle stratum k, the matched time period t, a suitable time window [Lbt , U bt ] could be
upstream vehicle i∗ = {i ∈ S|xij = 1} and the associated easily obtained. Assume that the travel time follows normal
matching probability is Pi∗ j . One naive approach to rule out distribution N (μt , σt ), then the tolerance interval with a 95%
those mismatches would be to impose a threshold value on confidence level can be utilized to define the time window, i.e.,
the matching probability. If Pi∗ j is greater than the threshold
value, then the travel time data regarding this vehicle j would [Lbt , U bt ] = [μt − 1.96σt , μt + 1.96σt ]. (22)
be retained for the following stratified sampling [see (19)].
However, in practical implementation, we find that the single Given the coefficient of variation φ, the time-window constraint
matching probability cannot really reflect the correctness of the can be rewritten as
matching. It is quite possible that the other matching probability [Lbt , U bt ] = [(1 − 1.96φ)μt , (1 + 1.96φ)μt ] . (23)
Plj ≈ Pi∗ j , l ∈ S; l = i∗ , which means that the VRI system
cannot distinguish between the candidate vehicles. To account Moreover, the prior knowledge can be approximated by the
for this problem, this study proposes a new measure to represent normal distribution N (μt , φμt ). Therefore, both of these two
the distinctiveness of the vehicle. For this specific vehicle j, the critical inputs of basic VRI can be derived from the prediction
distinctiveness value is defined as of the mean travel time in time period t. In other words, the
Pi∗ j self-adjusting of the time window and the prior knowledge can
(2)
(2)
, Pj is the second largest matching probability be completed by iteratively predicting the mean travel time for
Pj each time period. In this research, the self-adjusting of the time
of {Pij |∀i ∈ S}. (20) window for real-time application involves two major steps as
follows.
Recall that the proposed bipartite matching method finds the
• Interperiod adjusting: Based on current traffic information
matching result with the overall maximum probability [see
(e.g., average spot speed) and the mean travel time value
(14)]. In this case, for a certain downstream vehicle, it may
in a previous time period (i.e., obtained from the VRI
not be matched to the upstream vehicle with the maximum
system), one may predict the mean travel time value for the
likelihood. Therefore, by calculating the distinctiveness value,
next time period (i.e., interperiod adjusting), from which
one may get more information about the matching result, i.e.,
the time window is derived. The exponential forecasting
⎧
⎪ ≥ 1, vehicle j matches upstream vehicle with technique integrated with the average spot speed informa-
⎪
⎪
Pi∗ j ⎨ maximum probability tion is adopted during the interperiod adjusting process.
= = 1, vehicle j matches upstream vehicle with • Intraperiod adjusting: Since the nonrecurrent traffic con-
Pj
(2) ⎪
⎪
⎪
⎩ the second largest probability gestion (e.g., caused by incidents) is not predictable, the
< 1, vehicle j matches the other ones. additional intraperiod adjustment is required for providing
546 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

an appropriate time window under these extreme circum-

stances. An iterative bipartite matching method is pro-
posed for adjusting the time window, in which the basic
VRI is iteratively solved.
Note that the purpose of predicting the mean travel time is
to derive an appropriate time window, and the accuracy of the
prediction is not our major concern. As a matter of fact, the
estimated mean travel time is obtained from the VRI system
with the postprocessing technique.

A. Interperiod Adjusting
We introduce time series theory for short-term travel time
prediction. As a classical statistical approach, time series fore-
casting has already been evaluated with several other applica-
tions in transportation, such as short-term traffic flow prediction
[28] and traffic speed forecasting [29].
In this paper, the underlying model equation for the mean
travel time data is assumed as

μt = μ∗t + εt (24)

where μt is the mean travel time calculated from the VRI

system, μ∗t represents the ground truth data, and εt is the white Fig. 3. Detailed implementation of the improved VRI system.
noise error term. Our goal is to roughly forecast the mean travel
accuracy of the VRI system and, hence, unreliable travel time
time in period t + 1 (i.e., short-term prediction). Therefore, the
estimates. Thus, the additional intraperiod adjusting method
exponential smoothing technique integrated with spot speed
should be developed.
information is employed for this particular purpose.
1) Exponential Smoothing Technique: The smoothing (fore-
casting) equation is given as B. Intraperiod Adjusting

μ̌t+1 = μ̃t + ϕ(μt − μ̃t ) (25) From time period t + 1, the predicted time window
[Lbt+1 , U bt+1 ] and μ̃t+1 are derived during the interperiod
VtU + VtD adjusting process and then fed into the basic VRI system, from
μ̃t+1 = U +VD
μ̌t+1 (26)
Vt+1 t+1 which the raw travel time data can be obtained. By performing
the postprocessing technique (e.g., thresholding and stratified
where μ̃t+1 and μ̃t denote the forecasters of the mean travel (1)
sampling), an improved mean travel time estimator, i.e., μ̂t+1 ,
time in time period t + 1 and period t, respectively; ϕ rep-
is then calculated. In practice, the initial prediction of the time
resents the smoothing parameter that is calibrated from the
window may not be reliable (particularly when an incident hap-
historical data; VtU and VtD are the average speed at upstream
pens), which could significantly decrease the performance of
and downstream stations, respectively, during time period t; and (1)
U D the VRI system. Thus, it is expected that μ̂t+1 would not be ac-
likewise, Vt+1 and Vt+1 are the average speed at upstream and
curate. In light of this, an iterative process is devised to solve the
downstream stations, respectively, during time period t + 1.
basic VRI problems iteratively with different exogenous inputs
Equation (25) serves as a simple exponential estimation based
(i.e., time window and prior probability). To be more specific, a
on the estimates from previous steps, whereas (26) is a cor- (1) (1)
rection step by utilizing the average spot speed. The rationale new time window, i.e., [(1 − 1.96φ)μ̂t+1 , (1 + 1.96φ)μ̂t+1 ], is
(1)
behind (26) is as follows. If the average spot speeds at both calculated based on the estimated mean travel time μ̂t+1 . Then,
stations (i.e., upstream and downstream) decrease from period the basic VRI and the associated postprocessing technique are
t to period t + 1, the mean travel time during time period t + 1 performed again using this new time window. This iterative
is expected to be larger. process will continue until the relative change in the estimated
Following the prediction of μ̃t+1 , the time window for period mean travel time is sufficiently small. In this research, the error
t + 1 is given by for stopping tolerance of the convergence is given by
 
 (n) (n−1) 
[Lbt+1 , U bt+1 ] = [(1 − 1.96φ)μ̃t+1 , (1 + 1.96φ)μ̃t+1 ] . (27) μ̂t+1 − μ̂t+1  ≤ τ (28)

Based on these recursive formulas, one may be able to predict where superscript n represents the iteration number, and τ is
the mean travel time and the time window from period to the stopping tolerance (see Fig. 3).
period (i.e., interperiod adjusting). However, it should be noted To sum up, interperiod adjusting is designed to capture the
that a “bad” prediction could potentially lead to low matching traffic dynamics from period to period, whereas intraperiod
WANG et al.: VRI WITH SELF-ADAPTIVE TIME WINDOWS FOR REAL-TIME TRAVEL TIME ESTIMATION 547

Fig. 4. Test site in Bangkok, Thailand.

adjusting (i.e., iterative process) is introduced to handle the newly created vehicle records are then fed into the improved
nonpredictable traffic conditions (e.g., traffic incidents and VRI system.
bottleneck effect). To sum up, we simulate all traffic conditions (recurrent and
nonrecurrent traffic congestion) using VISSIM and implement
the proposed method in MATLAB. To be more specific, the
V. E XPERIMENTAL R ESULTS
experiments are performed under Windows 7 Home Premium
To verify the effectiveness and feasibility of the proposed and MATLAB v7.14 (R2012a) running on a Dell desktop with
improved VRI system, various simulation-based experiments an Intel(R) Core(TM) i3 CPU at 3.20 GHz and with 4.00 GB of
are conducted. In this research, a VISSIM-based simulation memory. It is easily observed that the computational time of the
model is devised to simulate freeway system operations under proposed method largely depends on the number of intraperiod
traffic demand and supply uncertainty (e.g., free flow, conges- iteration steps and the CPU time of the basic VRI system
tion, bottleneck effect, and traffic incident). (see Fig. 3). Some preliminary experiments also show that the
average CPU time used by the bipartite matching method in
A. Simulation Model Configuration and Calibration basic VRI under the free-flow condition is 0.0896 s, whereas
the average CPU time under the congested condition is about
Before presenting the experimental results, the detailed pro- 0.3294 s. Therefore, it is reasonable to believe that the improved
cedures for simulation model development and calibration are VRI system can be implemented for real-time application.
introduced. The test site for this research is a 34.9-km-long
three-lane freeway system in Bangkok, Thailand (see Fig. 4). At
B. Preliminary Comparison Between Basic VRI and
each station, a gantry-mounted video camera, which is viewed
Improved VRI
in the upstream direction, is installed, and the associated video
records are collected. Two segments are chosen for simulation To conduct the comparison of the basic VRI and the improved
model development: 1) a 3.6-km-long closed segment (i.e., VRI system, a VISSIM-based simulation model is designed for
between 08A and 10A, the green section in Fig. 4); 2) a 4.2-km- the closed segment between 08A and 10A (Westbound). During
long corridor with on/off-ramps (i.e., between 02A and 04A). the 4-h simulation time period, approximately 16 000 pairs of
The simulation model is then configured based on the exact vehicle records are generated. These vehicles can be roughly
roadway geometric feature, including the length of the segment, categorized into two types (see Section III): 70% of small
the location of on/off-ramps, and the number of lanes. cars and 30% of long vehicles. For this specific segment, the
To guarantee realistic representations of the simulated exper- associated image database, which includes 6280 pairs of vehicle
iments, model calibration is required. With the video records images, is built up. Therefore, a complete record for vehicle
collected at the test site, the individual vehicles can be detected i can be denoted as (IDi , ti , vi , Xi ), where IDi is the unique
and manually reidentified across multiple stations. Accordingly, identity derived from the simulation model; ti and vi are,
the ground truth data, such as vehicle counts, traffic demand, respectively, the arrival time and spot speed of vehicle i; and
and travel time data, can be obtained for model calibration. Xi represents vehicle feature data extracted from the vehicle
The correctly matched pairs of vehicle images are stored in the image. Based on these simulation data, the proposed VRI
image database for further application. system is performed and evaluated in terms of the matching
Upon completion of the simulation model configuration and accuracy and the effectiveness of mean travel time estimation.
calibration, the travel behavior and characteristic of each in- For the closed freeway segment, each vehicle can be detected
dividual vehicle (e.g., speed, vehicle type, and arrival time at at both stations. Therefore, it is expected that the matching
each station) can be collected. As the very heart of the proposed accuracy should be relatively higher, particularly for a static
method is the vision-based VRI, a vehicle image, which is time period (i.e., 5-min interval). However, for real-time appli-
randomly selected from the image database, is assigned to the cation, the potential changes in traffic conditions would lead
vehicle records generated from the simulation model. These to dramatic decrease/increase in matching accuracy. Fig. 5
548 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

The vehicle inputs are chosen such that all the traffic states
ranging from free-flow to congested can be activated. The
freeway segment operates under the free-flow condition in the
first stage (i.e., first hour). Congestion may be observed when
the vehicle inputs switch to the second stage. Then, the traffic
tends to a steady state of congestion during the following two
hours. In the fourth stage, congestion dissolve will be observed,
and the traffic will gradually be cleared from the freeway
system. Table I shows descriptive statistics for the outputs from
the VISSIM-based simulation model.
To validate the overall performance of the improved VRI
Fig. 5. Effectiveness of the self-adaptive time window. system, we run the method 50 times based on the simulation
outputs. For each run, the vehicle image is randomly selected
from the database and assigned to the vehicle records generated
from the simulation model. The root mean square error (RMSE)
and the mean absolute percentage error (MAPE) are applied as
performance indexes. The equation of the RMSE is given by

 I  2
1  50 
μi(s) − μ∗i
RMSE =  (30)
50 s=1 i=1 I

where μi(s) is the estimate for the ith time period and the
sth run, I indicates the total number of time periods, and μ∗i
Fig. 6. Effectiveness of the postprocessing technique under a fixed time represents the ith ground truth data. The MAPE is calculated as
window.
50 I   
1    μi(s) − μ∗i 
 × 100 .
shows the effectiveness of the basic VRI by employing the MAPE = (31)
self-adaptive time window. As the traffic volume significantly 50 × I s=1 i=1  μ∗i
increases during the second hour of the simulation experiment,
the corresponding traffic condition changes from free-flow to By simple calculation, the RMSE and the MAPE of the im-
congested. It is quite obvious that the fixed time window proved VRI system for a 5-min aggregation interval are 3.28 s
cannot handle this complicated situation (i.e., significant drop and 1.0%, respectively, whereas the RMSE of the basic VRI
in matching accuracy from period 12 to 26), whereas the VRI is 14.61 s. It is observed that the improved VRI clearly out-
system with a flexible time window can maintain a relatively performs the basic VRI. Fig. 7 shows the mean travel time
stable matching accuracy (around 60% of matching accuracy). estimates from one experiment. By integrating the average
On the other hand, given the fixed time window, a proper spot speed information (see Table I), interperiod adjusting can
postprocessing technique (i.e., thresholding and stratified sam- capture the traffic dynamics well, which could contribute to the
pling) can still improve the model performance from the travel following intraperiod adjusting (i.e., less intraperiod iteration
time estimation purpose. As shown in Fig. 6, the performance of steps).
the basic VRI significantly decays in the congested case due to
the decrease in matching accuracy. However, it is worth notic-
ing that the accuracy in mean travel time estimation improves a D. Performance Evaluation Under Bottleneck Effect
lot by imposing the postprocessing technique. As one of the major causes for freeway traffic congestion,
the freeway bottleneck can arise from many conditions, such as
high merging and diverging demand at on/off-ramps and lane
C. Performance Evaluation Under Recurrent
drops. In this part, we will evaluate the performance of the
Traffic Congestion
proposed method under the bottleneck effect. A 4.6-km-long
To further evaluate the performance of the improved VRI three-lane freeway segment between 02A and 04A (see Fig. 4)
system (i.e., with postprocessing and self-adaptive time win- is chosen as the test site. As shown in Fig. 8, one two-lane on-
dow) under recurrent traffic congestion (due to exceeding traffic ramp (2 km away from the upstream station) and one two-lane
demand), the three-lane closed segment (between 08A and off-ramp are distributed along this segment. We will ignore the
10A) is chosen for the test site. The stochastic vehicle inputs off-ramp at this location since it does not affect the bottleneck
of the VISSIM-based simulation model are defined as area. The vehicle inputs at the upstream station are the same as
⎧ (29), and we assume the following distribution of vehicle flows.
⎪ 4000 veh/h, 0 ≤ t ≤ 60 min
⎨
7000 veh/h, 60 ≤ t ≤ 120 min • 02A to 04A: 100%.
Q= (29)
⎩ 8000 veh/h, 120 ≤ t ≤ 180 min
⎪ • Off-ramp: 15% of the vehicle flow will exit from the two-
4000 veh/h, 180 ≤ t ≤ 240 min. lane off-ramp.
WANG et al.: VRI WITH SELF-ADAPTIVE TIME WINDOWS FOR REAL-TIME TRAVEL TIME ESTIMATION 549

TABLE I
D ESCRIPTIVE S TATISTICS F ROM S IMULATION O UTPUTS

Fig. 9. Performance of the proposed method (on-ramp vehicle flow distribu-

Fig. 7. Performance of the proposed method with exceeding traffic demand. tion: 25%).

TABLE II
P ERFORMANCE OF THE P ROPOSED M ETHOD W ITH D IFFERENT
V EHICLE F LOW D ISTRIBUTIONS

ber of intraperiod iteration steps would increase accordingly

(see Fig. 9).
Table II also shows the performance of the improved VRI
system under different vehicle flow distributions (i.e., on-ramp
vehicle flow distribution). Since the test site is a freeway
corridor with on/off-ramps, the vehicle arrives at the upstream
station may not necessarily appear at downstream. In addition,
some vehicles may enter this corridor through the on-ramp.
Thus, it is expected that the matching accuracy of the proposed
method is relatively lower. With the increase in the vehicle
flow from on-ramp, the performance would gradually decay.
However, it should be noted that the proposed method can still
Fig. 8. Merge and diverge along the freeway segment. perform well against the bottleneck effect.

• On-ramp to the freeway segment: 25% of the vehicle flow

will enter the freeway system through on-ramp (different E. Performance Evaluation Under Nonrecurrent
flow distributions will be tested in the experiment). Traffic Congestion
With respect to the above simulation outputs, we run the As the nonrecurrent congestion is largely produced by traffic
proposed method 50 times. Fig. 9 shows the estimation results incidents, this research will investigate the performance of the
from one experiment. Compared with the basic VRI, the pro- proposed method under traffic incidents. The test site is also a
posed method provides more reliable estimates of the mean three-lane closed segment (between 08A and 10A) and with the
travel time. In general, the bottleneck effect cannot be detected same vehicle inputs as (29). To mimic the situation of incident
through the average speed at upstream and downstream sta- happening, a parking lot located at lane 1 (2 km away from
tions, which means that interperiod adjusting cannot capture the upstream station) is utilized to simulate the incident vehicle
the traffic dynamics (congestion) well. Therefore, the num- (see Fig. 10). When an incident happens (i.e., incident starts
550 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

Fig. 12. Adjustment of the time-window constraint.

Fig. 10. Incident simulation.
TABLE III
P ERFORMANCE OF THE P ROPOSED M ETHOD W ITH
D IFFERENT I NCIDENT D URATIONS

travel time dramatically increases (periods 21 and 22). The

“abnormal” vehicle (seriously delayed by the incident at lane
1) and those normal vehicles may arrive at downstream during
the same time period. In this case, it renders a heavy burden
on the processing of the improved VRI system (e.g., larger
time window, more candidate vehicles, and of course, low
matching accuracy). Therefore, the results shown in Table III
Fig. 11. Performance of the proposed method (incident duration: 10 min;
starts from 90 min). are reasonable.

from 90 min), a vehicle would stop in the parking lot, and the
VI. C ONCLUSION AND F UTURE W ORKS
partial route is activated to simulate the driving behavior under
the incident condition. This paper aims to develop an improved VRI system based
The proposed algorithm is further tested with different in- on the authors’ previous work to estimate the real-time travel
cident durations (e.g., 10, 15, 20, and 30 min). Due to the time under traffic demand and supply uncertainty. A self-
unpredictability of the traffic incidents, interperiod adjusting adaptive time-window component is introduced into the ba-
cannot generate a suitable time window. Therefore, it is ex- sic VRI system to improve its adaptability against potential
pected that the steps of intraperiod iteration would significantly significant changes in traffic conditions. In addition, the as-
increase, particularly when a traffic incident occurs. Fig. 11 sociated postprocessing technique (i.e., thresholding based on
shows the mean travel time estimates from one experiment the matching probability and stratified sampling based on the
when the incident duration is 10 min. It is observed that the vehicle type) is employed to identify and rule out the erroneous
mean travel time sharply increases during time period 20 (i.e., travel time data. The proposed method is evaluated by conduct-
from 100 to 105 min). Hence, accordingly, the number of ing various representative simulation tests. Some performance
intraperiod iteration steps during this time period significantly indexes such as RMSE and MAPE are also introduced to
increases. Fig. 12 also illustrates the adjustment of the time- quantify the performance of this method.
window constraint for each iteration step. On the other hand, Further research will be focused on the real-world applica-
the basic VRI system cannot adapt well to the sudden changes tion of this proposed method. It is undeniable that the VIP
in traffic condition when an incident happens (see Fig. 11). Due systems are subject to the effects of inclement weather (e.g.,
to the fixed time-window constraint, the matching accuracy of rain and snow) and illumination changes. Under these cir-
basic VRI drops to 0% during time period 20, which eventually cumstances, the quality of the video image will dramatically
leads to a totally unreliable estimate of the mean travel time. decrease and, hence, compromise the effectiveness of vehicle
Parallel to the previous experiments, we also run the method feature extraction. During evening hours, the vehicle may still
50 times based on the simulation outputs for different incident be partially identified by detecting the vehicle headlight and
durations (e.g., 10, 20, and 30 min). The detailed estimation taillight. However, the color information and vehicle type may
results are shown in Table III. With the increase in incident not be obtained from the image. In this case, improving the
durations, the RMSE and the MAPE increase as well. With external lighting condition at each station may be a promising
the time propagation, it is also observed that the variance of way for vehicle feature extraction.
WANG et al.: VRI WITH SELF-ADAPTIVE TIME WINDOWS FOR REAL-TIME TRAVEL TIME ESTIMATION 551

As validated by the simulated tests, the proposed VRI system [19] K. Kwong, R. Kavaler, R. Rajagopal, and P. Varaiya, “Real-time measure-
for travel time estimation performs well under different sce- ment of link vehicle count and travel time in a road network,” IEEE Trans.
Intell. Transp. Syst., vol. 11, no. 4, pp. 814–825, Dec. 2010.
narios (e.g., recurrent traffic congestion, freeway bottlenecks, [20] A. Sumalee, J. Wang, K. Jedwanna, and S. Suwansawat, “Probabilistic
and minor traffic accidents). However, it is noteworthy that the fusion of vehicle features for re-identification and travel time estimation
proposed method may not work well under extremely abnormal using video image data,” Transp. Res. Rec., vol. 2308, pp. 73–82, 2012.
[21] C. Mallikarjuna, A. Phanindra, and K. R. Rao, “Traffic data collection
traffic conditions (e.g., severe traffic accident with longer inci- under mixed traffic conditions using video image processing,” J. Transp.
dent duration). As explained in Section V-E, the longer incident Eng., vol. 135, no. 4, pp. 174–182, Apr. 2009.
duration would inevitably lead to a larger time window, low [22] L. A. Klein, M. K. Mills, and D. R. P. Gibson, Traffic Detector Handbook,
3rd ed. Washington, DC, USA: Federal Highway Administration, 2006.
matching accuracy of the proposed VRI system, and hence an [23] W. Lin and D. Tong, “Vehicle re-identification with dynamic time win-
unreliable travel time estimator. Therefore, future efforts should dows for vehicle passage time estimation,” IEEE Trans. Intell. Transp.
be dedicated to overcome these drawbacks. As the lane block- Syst., vol. 12, no. 4, pp. 1057–1063, Dec. 2011.
[24] M. Ndoye, V. F. Totten, J. V. Krogmeier, and D. M. Bullock, “Sensing and
ing caused by incidents would produce a significant impact on signal processing for vehicle reidentification and travel time estimation,”
the travel time experienced by the vehicles at different lanes, IEEE Trans. Intell. Transp. Syst., vol. 12, no. 1, pp. 119–131, Mar. 2011.
one possible way is then to perform stratified sampling based [25] A. T. G. Thiang and R. Lim, “Type of vehicle recognition using template
matching method,” in Proc. Int. Conf. Electr., Electron., Commun., Inf.,
on the vehicle lane position. 2001, pp. 1–5.
[26] A. Bhattacharyya, “On a measure of divergence between two statistical
populations defined by their probability distributions,” Bull. Calcutta
R EFERENCES Math. Soc., vol. 35, no. 1, pp. 99–109, 1943.
[1] D. Ni and H. Wang, “Trajectory reconstruction for travel time estimation,” [27] B. R. Hellinga and L. Fu, “Reducing bias in probe-based arterial link
J. Intell. Transp. Syst., vol. 12, no. 3, pp. 113–125, Aug. 2008. travel time estimates,” Transp. Res. Part C, vol. 10, no. 4, pp. 257–273,
[2] F. Soriguera and F. Robuste, “Requiem for freeway travel time estimation Aug. 2002.
methods based on blind speed interpolations between point measure- [28] M. C. Tan, S. C. Wong, J. M. Xu, Z. R. Guan, and P. Zhang, “An aggre-
ments,” IEEE Trans. Intell. Transp. Syst., vol. 12, no. 1, pp. 291–297, gation approach to short-term traffic flow prediction,” IEEE Trans. Intell.
Mar. 2011. Transp. Syst., vol. 10, no. 1, pp. 60–69, Mar. 2009.
[3] H. B. Celikoglu, “Flow-based freeway travel-time estimation: A compara- [29] Q. Ye, W. Y. Szeto, and S. C. Wong, “Short-term traffic speed forecasting
tive evaluation within dynamic path loading,” IEEE Trans. Intell. Transp. based on data recorded at irregular intervals,” IEEE Trans. Intell. Transp.
Syst., vol. 14, no. 2, pp. 772–781, Jun. 2013. Syst., vol. 13, no. 4, pp. 1727–1737, Dec. 2012.
[4] C. Lindveld, R. Thijs, P. Bovy, and N. V. der Zijpp, “Evaluation of online
travel time estimators and predictors,” Transp. Res. Rec., vol. 1719, no. 1,
pp. 45–53, 2000.
[5] R. Li, G. Rose, and M. Sarvi, “Evaluation of speed-based travel time
estimation models,” J. Transp. Eng., vol. 132, no. 7, pp. 540–547,
Jul. 2006.
[6] S. M. Quayle, P. Koonce, D. DePencier, and D. M. Bullock, “Arterial per-
formance measures with media access control readers: Portland, Oregon,
pilot study,” Transp. Res. Rec., vol. 2192, pp. 185–193, Dec. 2010.
[7] A. Hofleitner, R. Herring, P. Abbeel, and A. Bayen, “Learning the dynam-
ics of arterial traffic from probe data using a dynamic Bayesian network,”
IEEE Trans. Intell. Transp. Syst., vol. 13, no. 4, pp. 1679–1693, Dec. 2012.
Jiankai Wang received the B.S. and M.S. degrees
[8] S. Chang, L. Chen, Y. Chung, and S. Chen, “Automatic license plate
in mathematics, with a focus on operations re-
recognition,” IEEE Trans. Intell. Transp. Syst., vol. 5, no. 1, pp. 42–53,
search, applied optimization, and image reconstruc-
Mar. 2004.
tion, from Nanjing University, Nanjing, China, in
[9] G. Rose, “Mobile phones as traffic probes: Practices, prospects and is-
2007 and 2010, respectively. He is currently working
sues,” Transp. Rev., vol. 26, no. 3, pp. 275–291, May 2006.
toward the Ph.D. degree in transportation engineer-
[10] C. Sun and S. G. Ritchie, “Individual vehicle speed estimation using single
ing in the Department of Civil and Environmental
loop inductive waveforms,” J. Transp. Eng., vol. 125, no. 6, pp. 531–538,
Engineering, The Hong Kong Polytechnic Univer-
Nov./Dec. 1999.
sity, Kowloon, Hong Kong.
[11] B. Coifman, “Estimating travel times and vehicle trajectories on freeways
His research interests include image processing,
using dual loop detectors,” Transp. Res. Part A, vol. 36, no. 4, pp. 351–
traffic surveillance and management, and optimiza-
364, May 2002.
tion in intelligent transportation systems.
[12] B. Coifman and S. Krishnamurthy, “Vehicle reidentification and travel
time measurement across freeway junctions using the existing detec-
tor infrastructure,” Transp. Res. Part C, vol. 15, no. 3, pp. 135–153,
Jun. 2007.
[13] S. Kamijo, T. Kawahara, and M. Sakauchi, “Vehicle sequence image
matching for travel time measurement between intersections,” in Proc.
IEEE Int. Conf. Syst., Man, Cybern., 2005, vol. 2, pp. 1359–1364.
[14] C. C. Sun, G. S. Arr, R. P. Ramachandran, and S. G. Ritchie, “Vehicle
reidentification using multidetector fusion,” IEEE Trans. Intell. Transp.
Syst., vol. 5, no. 3, pp. 155–164, Sep. 2004.
[15] D. P. Watling and M. J. Maher, “A statistical procedure for estimating a
mean origin–destination matrix from a partial registration plate survey,”
Transp. Res. Part B, vol. 26, no. 3, pp. 171–193, Jun. 1992. Nakorn Indra-Payoong received the Ph.D. degree
[16] D. P. Watling, “Maximum likelihood estimation of an origin–destination in computing from the University of Leeds, Leeds,
matrix from a partial registration plate survey,” Transp. Res. Part B, U.K.
vol. 28, no. 4, pp. 289–314, Aug. 1994. He is currently an Associate Professor of logistics
[17] B. Coifman, “Vehicle re-identification and travel time measurement in and supply chain management with the Faculty of
real-time on freeways using existing loop detector infrastructure,” Transp. Logistics, Burapha University, Bangkok, Thailand.
Res. Rec., vol. 1643, no. 1, pp. 181–191, 1998. His research interests include large-scale network
[18] K. Kwong, R. Kavaler, R. Rajagopal, and P. Varaiya, “Arterial travel optimization and intelligent transport systems.
time estimation based on vehicle re-identification using wireless magnetic
sensors,” Transp. Res. Part C, vol. 17, no. 6, pp. 586–606, Dec. 2009.
552 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 15, NO. 2, APRIL 2014

Agachai Sumalee received the B.Eng. degree in civil Sakda Panwai received the B.Sc.Id.Ed. degree
engineering from King Mongkut’s Institute of Tech- in civil engineering (first-class honor) from King
nology Ladkrabang, Bangkok, Thailand, in 1999 and Mongkut’s Institute of Technology Thonburi,
the M.Sc.(Eng.) and Ph.D. degrees in transportation Bangkok, Thailand, in 1996; the M.E. degree in
planning and engineering from the University of transportation engineering from the Asian Institute
Leeds, Leeds, U.K., in 2000 and 2004, respectively. of Technology, Pathumthani, Thailand, in 1998;
He is an Associate Professor with the Department and the Ph.D. degree from The University of
of Civil and Environmental Engineering, The Hong Queensland, Brisbane, Australia, in 2007.
Kong Polytechnic University, Kowloon, Hong Kong. For over 15 years, he has been with the Express-
He has published more than 60 research papers in way Authority of Thailand (EXAT), Bangkok, where
top peer-reviewed journals. His research interests he is responsible for conducting feasibility studies,
include transit planning, intelligent transport systems, network modeling and highway design, research studies (e.g., traffic impact analysis and traffic stud-
optimization, and transport economics. ies), and the development of intelligent transportation system (ITS) applications
Dr. Sumalee is the Editor-in-Chief of Transportmetica B: Transport Dynam- (e.g., incident detection system, travel time estimation, and road and vehicle
ics, the Editor of Transport Policy, an Associate Editor of Networks and Spatial emission including the urban heat island study). His research interests include
Economics, and an Editorial Board Member of several prestigious journals. He ITS applications, smart infrastructure, and advanced ITS system modeling
has received several prizes and awards for his research works. using a traffic simulation technique.
Dr. Panwai received an EXAT Scholarship for his Master’s degree in 1996,
an International Road Federation Fellowship award in 2002/2003, a Royal Gov-
ernment (RTG) Scholarship for his doctoral study in 2003, and the European
Marie Curie Award for attending a training course in Berlin, Germany, in 2005.
In 2011, he was invited to Hiroshima University for a research fellowship
program.