Joint International Mechanical, Electronic and Information Technology Conference (JIMET 2015)

A Robust Framework for Vehicle Detection and Tracking Based on

Particle Filter
Huihui Liu1,a, Yong Liu2,b and Haiqing Du3,c
1
Beijing University of Posts and Telecommunications, Beijing, China
a
[email protected], [email protected], [email protected]

Keywords: Vehicle detection/tracking; Foreground extraction; Particle filter; Histogram of oriented

gradients; HSV color space; Occlusion segmentation;

Abstract. Vehicle detection and tracking have been promising application in traffic surveillance and
vehicular network. However, the vision-based approach still remains a challenging task due to the
problems of illumination variation, shadow and occlusion. In this paper, we propose a robust
framework mainly concatenates on two aspects: adaptive vehicle detection with shadow removal, and
vehicle tracking with occlusion handling. Firstly, for vehicle detection stage, an improved ViBe
algorithm with ghost suppression is adopted to extract moving vehicle region. Then moving shadow
is removed by integrating improved color with texture feature. Then, aiming to achieve the
multi-vehicle tracking, we propose an enhanced histogram of the oriented gradient combined with
HSV color space based on particle filter (ECHOGPF). Finally, we employ the occlusion detection
and occlusion segmentation to refine our system, which are based on one-dimensional maximum
entropy and the least square ellipse fitting. Experiments on popular datasets show that our proposed
system has a good effectiveness, e.g., the accuracy is 95% on the vehicle tracking.

Introduction
The vehicle detection and tracking play a vital role in many vision-based intelligent transportation
system applications. As existing gradual or sudden illumination changes, occurrence of shadow and
vehicle occlusion, there are great difficulties among vehicle detection and tracking. We aim to design
a robust framework, which could handle these problems. Our proposed system is consisted of
background modeling, particle filter, one-dimensional maximum entropy and features like HSV color
space, local texture, and histogram of oriented gradient.
Background modeling [1, 2] is the most commonly method used in the task of real-time detection.
Compared with the GMM algorithm [4] and codebook algorithm [5], the visual background extractor
[3] has got more attention due to its high precision, and fast processing speed. Li et al. [6] propose an
improved ViBe algorithm that could suppress ghost well. As the shadow could cause objects merging,
the distortion of shape and tracking loss, Cucchiara R et al. [7] use HSV color space, and Javed and
Shah [8] apply the gradient direction to detect the shadow. However, there exists some shortcomings
for the methods, which are based on HSV color space, i.e., some parameters should be adjusted
depending on different traffic environments and single feature can’t always work well. Therefore, we
integrate the local texture feature with an improved non-parametric feature of HSV color space to
remove shadow.
After detected the interest objects, we present a new appearance observation likelihood model in
which the similarity of appearance between targets with a template is calculated. Feature is
indispensable for appearance modeling. Sugimura et al. [9] uses color histogram for crowd tracking.
Dalal et al. [10] apply histogram of oriented gradient in human detection. We combine enhanced the
HOG feature with color histogram for precision and complementary description of feature. A good
appearance model is of great aid to the task of tracking.
As the occlusion could result large tracking error, it has been one of the main research topics in
video target tracking field for a long time. In general, occlusion is segmented by three kinds of
methods: feature-based [12], model-based [13], and contour-based [14]. The feature-based approach
usually costs too much time and makes vehicle tracking hard to be real-time. Kamijo et al. [13] prop-

© 2015. The authors - Published by Atlantis Press 803

 Fig. 1. The framework for vehicle detection and tracking.

osed a solution based on random field models. However, the model-based method introduces
unknown variables and usually could not adapt to the changed environment. The contour-based
method could be failed if the foreground is incomplete. Comaniciu et al. [15] proposed a new object
appearance based on mean-shift for partial occlusion problems. Wu et al. [16] proposed a partial
occlusion detection approach based on the reduction of particle weights within the occluded regions.
Many related occlusion segmentation methods are based on specific tracking algorithm. In this paper,
we propose an adaptive occlusion segmentation method, which can handle multi-vehicle occlusion
and be applied to other tracking framework. Our framework is based on the well-known particle filter
[11], which is robust and flexible, and which has the ability to deal with the non-Gaussian and
nonlinear models. As shown in Fig. 1, our system consists of the following modules: vehicle
detection (ViBe background modeling, update and shadow removal), vehicle tracking (occlusion
detection and segmentation, ECHOG appearance model, and particle filter tracking), and access to
traffic information.

Vehicle Detection and Tracking Framework

Adaptive Background Modeling for Vehicle Detection. The algorithm of background modeling
ViBe (visual background extractor) introduces randomization. The algorithm mainly comprises of
three steps. The first step is to initialize the model using the strategy of random selection. The second
step is to update the background model. The third step is to randomly propagate the pixel of the
background image to its neighborhood based on the principle of pixel spatial relativity.

(a) (b)

(c) (d)
Fig. 2. The results of foreground extraction: (a) Frame 5 with ViBe detection, (b) Frame 35 with
ViBe detection, (c) Frame 150 with ViBe detection, (d) Frame 35 with improved ViBe detection.

How to initialize the background model is a significant aspect for vehicle detection. There exists
two ways to initialize in the previous method. The one method which uses the first frame of the video
sequence to initialize the background model and this method achieves fast speed of initialization.
However, the ghost will appear when the target exists in the first frame. In this case, the real

804
background pixel could not match the background model once the target left and this region becomes
ghost. Therefore, we adopt multi-frame initialization algorithm which is based on ViBe detection [6].
The improved algorithm can suppress ghost well. An intuitive description is shown in Fig. 2. Fig. 2(c)
shows the result for original ViBe algorithm, and Fig. 2(d) shows a better result which is obtained by
our method.

(a) (b) (c) (d)

Fig. 3. Example of shadow detection for foreground extraction. (a) The original frame. (b) Mask of
shadow using LBP feature. (c) Mask of shadow using improved HSV color space. (d) Foreground
image.

Moving Shadow Detection Integrating Improved Color and Texture. Using the background
modeling, each pixel could be classified into two classes: foreground and background. However, the
foreground image is likely to contain the shadow. It causes incorrect object information and may lead
to occlusion. Therefore, shadow detection is significant during tracking. Many previous studies have
been discussed about the shadow elimination. However, the existing models, which are based on
HSV color space or based on texture, i.e., LBP, all have some drawbacks. For example, if the color
between moving object with background is similar, the region of the object may be detected as
shadow mistakenly. For the most methods, the threshold parameters should be constantly optimized
[15]. Therefore, we combine the feature of LBP and improved non-parametric HSV color space. The
Fig. 3 is an intuitive description for the shadow detection.
In the region covered by shadow, the change of luminance is bigger than hue and saturation. The
mask of shadow C HSV is calculated by Eq. 1.
Iv (x,y)
1, if 0 ≤ Btv (x,y) ≤ 1 ⋀ |Ith (x, y) − It−1
h
(x, y)| ≤ 1
CtHSV (x, y) ={ t (1)
0, otherwise

where Itv (x, y) and Btv (x, y) are pixel value of luminance channel for the given foreground image and
background image of frame t. We utilize the Ith (x, y)between two adjacent frames to discriminate
shadow. The advantage of this method is that there are no parameters in the above process. Then, we
use LBP operators to compute the texture mask LLBP . Texture similarity ρ between the regions of
shadow with background is measured by histogram intersection using the Eq. 2.

ρ = ∑N−1 I B
n=0 min(hn , hn ) (2)

where hIn denotes histogram of current frame, and hBn denotes histogram of background image. Once
obtaining the likelihood, the foreground pixels could be classified into shadow and target.

1, 𝑖𝑓 𝜌 ≥ TLBP
LLBP
t (x, y) = { (3)
0, otherwise

where TLBP is set by the user according to actual video surveillance environment. LLBP is shown in
Fig. 3(b) and C HSV is shown in Fig. 3(c).To get the result of shadow segmentation, we need to AND
C HSV to LLBP . Certain areas of vehicles are detected as shadow mistakenly. So we refine the final mask
though connected-components analysis. Fig. 3(d) shows the result of foreground image.

805
ECHOGPF Vehicle Tracking Algorithm. The output of the vehicle detection is a binary object
mask which is used to perform tracking. Tracking is a consecutive task which aims to search the
best-match object in the current frame and update feature of the template for an accurate match in the
subsequent frame. Our proposed method，ECHOGPF, mainly consists of three parts. The first part is
color histogram and enhanced HOG (ECHOG). The second part is particle filter tracking over time
based on ECHOG appearance model. The third part is the appearance template matching used for
checking whether the candidate bounding box matches the target in previous frame or not. In this
paper, the template is progressively updated at each time instant.
ECHOG Appearance Model. Color histogram represents the distribution of colors in an image. It
is the most popular appearance model in multi-object tracking system because of its effectiveness on
obtaining statistical information. We compute a 3D HSV color histogram of concerned vehicles. One
point worth noting is that the number of bin should be balanced because of bad discriminant effect
when being too large or small. In our model, we choose 8 bins for h-channel, 12 bins for s-channel,
and 3 bins for v-channel. Therefore, the total number of elements of HSV color histogram descriptor,
which is used to compute similarity distance, is 288. However, the color histogram has a drawback
due to the loss of the spatial information. Therefore, we improve the appearance model through
combining the enhanced HOG descriptor [16].
Histogram of oriented gradient (HOG) is one kind of gradient descriptor, which is usually
employed to image processing and object detection. Firstly, color-image is preprocessed by graying.
We compute gradient by convolving gray image with a vertical kernel [−1,0,1]Tand a horizontal
kernel [−1,0,1]T. The size of all object image patches is resized to 32×32. Next，the image is divided
into small cells with the size of 8×8. The orientation of gradient ranged between 0o with 180o is
uniformly divided into nine bins. In each cell, we compute the magnitude of gray value in all bins. A
block is formed by 4 each close cells. So there are 36 bins in one block and nine blocks in one image.
The nine blocks forma 324-dimension feature vector that is used as a HOG descriptor. The RDHOG
descriptor used for enhancing the description of the object reflects difference between the central
block and its neighbor blocks. The difference is computed by the Eq. 4.

RDj (i) = Scenter (i) − Sj (i), j = 1,2 … … 9, j ≠ center (4)

where Scenter (i) is ith bin in the central block and Sj (i) is ith bin in jth block. RDj (i) is the ith bin of
RDHOG descriptor in jth block. In this way, 288 bins make up a 288-dimension feature vector. By
combining 3D HSV histogram with HOG and RDHOG descriptor, the model becomes more robust. It
can not only work well in normal light condition but also the low contrast environment. The above
descriptors are integrated into the particle filter framework for vehicle tracking.
Particle Filter Framework. Particle filter provides a probabilistic framework for recursive
dynamic state estimation [17]. The goal is to estimate the posterior distributions of state variables that
are represented as sets of weighted particles. A particle is one possible state of target. The filter is
initialized by the state of target at the start time. The steps are shown as follows. The first stage is
prediction. Particles are propagated from frame t-1 to frame t using the following motion model

S(i, t) = A × S(i, t − 1) + W(i, t − 1) (5)

where S(i, t) denotes the state of ith particle in frame t. The parameter A denotes state-transition
matrix and W denotes Gaussian distribution of random variables. The second stage is to update. We
compute a weight for each propagated particle using our appearance model and normalize all of
weights to sum to one. The third stage is to resample to avoid particle degeneracy phenomenon. The
particles with low weight are deleted and high weight particles are replicated. As a result, a new set of
equal weight particles can be generated [18].

806
 (a) (b) (c) (d)

(e) (f)
Fig. 4. (a) Frame 1224 of Video1. (b) An extracted target from frame 1224. (c) Target binary
foreground image. (d) Target HOG image. (e) - (f) Target HSV histogram.
j
Object Matching. Object matching aims to determine which object On in current frame is the
j
most relevant with the object On−1 in previous frame. Therefore, we need to measure the similarity
between the observation and template. The color or HOG feature is used to describe in many other
measurements. In this paper, we build a new combination evaluation for tracking vehicles. Firstly, we
resize the bounding box of the target to the same size of template. Fig. 4(c) shows the target binary
foreground image of target in Fig. 4(b). The HOG image and HSV histogram of the target are shown
t
in Fig. 4.The Bhattacharyya similarity f(ht|xj,i , hj) between model histogram and observed histogram
is calculated by the Eq. 6.

t j t
f(ht|xj,i , hj) ∝ e1−d(h ,h ) (6)

Where

j
d(hj , ht ) = ∑N−1 t
i=0 √hi hi (7)

j j
Using the HSV and HOG descriptors, we can get HSV weight dHSV,n and HOG weight dHOG,n of
j
jth particle in frame n. hi denotes ith bin in the observed histogram of the jth particle and hti denotes
the ith bin in the histogram of template of the jth particle. In this paper, we set N is 288 for HSV
j
descriptor and 324 for HOG descriptor. The RDHOG weight DRDHOG,n is calculated by Eq. 8

j j
DRDHOG,n = ∑D−1 t
i=0 |bi − bi | (8)

j
where bi and bti denotes the ith bin of RDHOG descriptor of the observed object and template. The
j
number of D is 288. DRDHOG,n denotes the difference between the central block and its neighbor
blocks. The weight of jth particle in frame n is given in Eq. 9.
j j j j
wn = (dHSV,n × dHOG,n )⁄DRDHOG,n (9)

The weight of particle reflects the degree of similarity between candidate bounding box with the
j j j
template. It will increase when dHSV,n or dHOG,n increases and decrease when DRDHOG,n increases.

807
 (a) (b)

(c) (d) (e) (f) (g) (h)

Fig. 5. The procedure of occlusion segmentation. (a) Vehicle tracking without handling occlusion. (b)
Vehicle tracking with handling occlusion (c) Occlusion detection and candidate region extraction. (d)
Histogram equalization. (e) Gray-scale transformation. (f) One-dimensional maximum segmentation.
(g) Ellipse fitting. (h) Line scanning and foreground segmentation.

A Novel Occlusion Handling Method. Due to the phenomenon of interaction and overlapping
between vehicles, the detected candidate bounding box may contain multiple vehicles. The
appearance of a vehicle will exhibit different in two adjacent frames when occlusion exits, and thus
make it difficult to associate the same vehicle extracted from the frames. The two main problems
should be solved in occlusion segmentation that are occlusion detection and multi-vehicle
segmentation.
A lot of study work has been done for occlusion detection. Our occlusion detection is based on
geometrical shape of area because the adaptable algorithm needs less prior knowledge and processing
speed is relatively fast. When occlusion happens, the duty ratio R d = R V /R B and aspect ratio R a =
R H /R w become abnormal where R V denotes the area of the vehicle region and R B denotes the
envelop rectangle. R H and R w are the height and width of the rectangle. The detailed detection is
designed as follows. The smaller value of R d , the more likely the region is abnormal and R a is on
the contrary. If R d <Yd or R a >YHW , we regard this region as abnormal. Under the basis of vast
analysis and careful studies, we set Yd as 0.75 and YHW as 1.55. In non-occlusion cases, we use
particle filter to predict the positions of vehicles in current frame. In occlusion cases, we must
segment all vehicles in one candidate box, and track them in particle filter framework.
To solve the problem of the segmentation of vehicles with occlusion, we first adopt the
one-dimensional maximum entropy method to split the image into a binary segment image, and some
result are shown in Fig. 5(f). In this paper, the candidate region is a rectangle surrounding the vehicles.
For better splitting, we apply histogram equalization and image gray-scale transformation to enhance
the contrast of candidate bounding box. Then the least square ellipse is utilized to fit the connected
region of the binary image. The long axis of ellipses is in the direction of lane. Lines that are parallel
to the major axis of these ellipses divide the foreground boundary into several sub-boundaries
depending on the number of ellipses. For the case of multi-vehicle occlusion, three or more ellipses
can be detected with the one-dimensional maximum entropy method. The lines scan the region
between the two each close centers of ellipses to find the least sum of distance from the centroids of
sub-boundaries (Fig. 5(g) and Fig. 5(h)). In contrast to many algorithms which mainly focus on
occlusion of two vehicles, three or more vehicles can be segment effectively. Fig. 5(b) shows the
result of segmentation and tracking for object#3 in Fig. 5(a).

808
Experiments and Results
Experiment setup. In this section, we evaluate our proposed system based on several video
sequences. These video sequences contain various traffic environments respectively, and we could
detail describe each video in the following. Vehicle video1 denotes low or medium traffic flow. The
highway II video denotes the medium traffic flow with the occurrence of shadow. Vehicle video2
denotes a high traffic flow with serious occlusion.
The experiments are performed in PC platform with Intel Core i5 with a 3.3-GHz central
processing unit and 3GB of RAM. Our approach is developed in C++ using OpenCV Library [19].
OpenCV is a library of programming functions mainly focuses on image processing and computer
vision. The obtained experiment results are compared with the other two similar methods proposed by
Lien et al. [20] and Scharcanski et al. [21]. The method proposed by Lien et al. [20] mainly focuses
on texture-based object segmentation and tracking. The method proposed by Scharcanski et al. [21]
is designed to detect occlusion and track vehicle using a Rayleigh distribution based on particle filter.
Performance analysis. For the evaluation of tracking performance, the measurement includes
three factors: correct tracking which denotes the number of correctly tracking frames, missing
tracking which denotes the number of missing tracking frames and false tracking which denotes the
number of false tracking frames.
Table 1 shows the comparison between our proposed method with the other two methods. Aim to
evaluating the effectiveness of our framework, we evaluate all of them on three videos. Although the
method of Lien could effectively detect vehicle without background model, the texture-based target
tracking method may fail when the texture feature of vehicles is similar. Compared with the method
of Lien, our appearance model could hold essential characteristic of vehicles more accurately. From
the table, we could see that the system of Lien has lower correct fraction in video2 due to it can’t deal
with occlusion. The method proposed by Scharcanski could handle occlusion and the correct fraction
is slightly higher. However, this algorithm has limitations in dealing with occlusion because it can’t
be applied to other tracking system. And it performs worse in poor lighting or shadow environments
because the method can’t detect shadow well. Our system has high correct fraction in traffic scene
with shadow which shows the robustness.
As expected, our approach outperforms the other two systems. Experimental results show that our
proposed framework was running at about 20 frames per second and the accuracy of tracking can
approach 95%. However, when the vehicles in the video are so far away that appear too small, their
local appearance characteristics and spatial relationships can’t be captured completely. That is the
reason of high missing tracking ratio and false tracking ratio. One solution is to set up the region of
interest in the central region of the video. In this way, we can discard the region near the border and
improve our system performance.

Table 1. Performance comparison of vehicle tracking

Correct Missing False Correct Error
Video Tracking Algorithm
Tracking Tracking Tracking Fraction(%） Fraction(%)
Our system 3146 55 49 96.8% 3.2%
Video1 Lien et al.[20] 2935 234 81 90.4% 9.6%
Scharcanski et al.[21] 3020 138 92 92.9% 7.1%
Our system 450 8 4 97.4% 2.6%
Highway
Lien et al.[20] 436 7 19 94.3% 5.7%
II
Scharcanski et al.[21] 424 25 13 91.7% 9.3%
Our system 280 8 12 93.3% 6.7%
Video 2
Lien et al.[20] 260 11 29 86.6% 13.4%

809
 Scharcanski et al.[21] 281 10 9 93.6% 7.4%

(a) Frame 636 (b) Frame 638 (c) Frame 640 (d) Frame 165 (e) Frame 171

(f) Frame 226 (g) Frame 20 (h) Frame 90 (i) Frame 97

Fig. 6. The tracking results of test sequences. The results of vehicle tracking for video1 are shown in
(a), (b) and (c) that present our proposed method is able to track multiple vehicles in real traffic scene.
The results for highway II video are shown in (d), (e) and (f) that present the results of real-time
shadow scene. The results for video2 are shown in (g), (h) and (i) that present the tracking results
before, during and after vehicles are occluded. It shows that our framework can handle occlusion
problem and track vehicles accurately.

Summary
In this paper, based on particle filter, we present a robust processing system for vehicle detection and
tracking while satisfying the real-time constraints, which can handle shadow and occlusion
effectively. The improved feature fusion is efficient for shadow detection. And the enhanced
appearance model works well in different environments. Our proposed occlusion segmentation can
handle occlusion accurately. Experiments under different illumination intensity and traffic flow
demonstrate the robustness and efficiency of our framework.

References
[1] Kamijo, S., Matsushita, Y., Ikeuchi, K., and Sakauchi, M.,“Traffic monitoring and accident
detection at intersections,” IEEE Transaction on Intelligent Transportation Systems, Vol.1, No.2, pp.
703-708(2000).
[2] Zhou, J., Gao, D., and Zhang, D., “Moving vehicle detection for automatic traffic monitoring,”
IEEE Transaction On Vehicular Technology, Vol.56, No.1, pp.51-59(2007).
[3] O.Barnich, M. Van Droogenbroeck. ViBe: A powerful random technique to estimate the
background in video sequences [C].Proceedings of IEEE International Conference on Acoustics,
Speech and Signal Processing, 945-948 (2009).
[4] C. Stauffer, W. Eric, L. Grimson. Learning patterns of activity using real-time tracking [J]. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 22(8): 747-757(2000).
[5] K. Kim, T. Chalidabhongse, D. Harwood, and L. Davis, “Real-time foreground-background
segmentation using codebook model,” Special Issue on Video Object Processing, Real-Time Imag.,
Vol. 11, pp.172-185 (2005).
[6] Yongqing Li, Wanzhong Chen, Rui Jiang. The integration adjacent frame difference of improved
ViBe for foreground object detection [C]. Proceedings of 7th International Conference on Wireless
Communications, Networking and Mobile Computing, pp. 1-4 (2011).

810
[7] Cucchiara R, Grana C, Piccardi M, et al. Improving shadow suppression in moving object
detection with HSV color information [C].Proceedings of IEEE Intelligent Transportation Systems,
pp. 334-339 (2001).
[8] JAVED O, SHAH M. Tracking and object classification for automated surveillance
[C].Proceedings of IEEE European Conference on Computer Vision. [S.I.], pp. 343-357 (2002).
[9] Daisuke Sugimura, Kris Makoto Kitani, Takahiro Okabe, Yoichi Sato, and Akihiro Sugimoto.
Using individuality to track individuals: clustering individual trajectories in crowds using local
appearance and frequency trait .In ICCV. 1467-1474(2009).
[10] N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In CCPR(2005).
[11] J. Scharcanski, A. B. de Oliveria, P.G. Cavalcanti, and Y.YARI, “A particle-filtering approach
for vehicular tracking adaptive to occlusions,” IEEE Transaction on Vehicular Technology, Vol.60,
No.2, pp.381-389 (2011).
[12] LINDERBERG T. Feature detection with automatic scale selection [J]. International Journal of
Computer Vison, 30(2):79-116 (1998).
[13] KAMIJO S, SAKAUGHI M. Segmentation of vehicles and pedestrians in traffic scene by
spatio-temporal Markov random field model [C]. Proceedings of the 21st International Conference on
Data Engineering Workshops, pp.1-8 (2005).
[14] ZHANG Xin, JI Xiu-hua. An improved Harris corner detection algorithm for noised images
[J].Advanced Materials Research, 433: 6151-6156(2012).
[15] D. Comaniciu, V. Rameshand, P. Meer, Real-time tracking of non-rigid objects using mean-shift,
in: Computer Vision and Pattern Recognition, pp. 142-149 (2000).
[16] WU B, KAO C, JEN C, et al. A relative discriminative histogram of oriented gradients-based
particle filter approach to vehicle occlusion handling and tracking [J]. IEEE Transactions on
Industrial Electronics, 61 (8): 4228-4237 (2014).
[17] M.S.Arulampalam, S.Maskell, N.Gordon, and T.Clapp.A tutorial on particle filters for online
nonlinear/non-gaussian Bayesian tracking. IEEE Transactions on Signal Processing, 50(2): 174-188
(2002).
[18] Rui, Y, Chen, Y.: Better proposal distributions: Object tracking using unscented particle filter. In:
IEEE Conference on Computer Vision and Pattern Recognition, Vol 2, pp.786-793(2001).
[19] OpenCV on SourceForge: http://sourceforge.net/projects/opencvlibrary/.
[20] C.C. Lien, Y.T. Tsai, M.II. Tsai, and L.G Jang, “Vehicle counting without background
modeling,”17th International Conference on Advances in Multimedia Modeling, vol. Part 1,
Springer-Verlag, Berlin, IIcidelberg, pp.446-456 (2011).
[21] Scharcanski, J.; de Oliveira, A.B.; Cavalcanti.P.G.; Yari, Y. (2011). “A Particle-Filtering
Approach for Vehicular Tracking Adaptive to Occlusions”, IEEE Transactions on Vehicular
Technology, Vol. 60, No.2, pp.381-389 (2011).

811