Model update particle filter for multiple objects detection and tracking
Yun Ji ZHAO,Hai Long PEI
Abstract— Multiple objects tracking is a challenging task.
This article presents an algorithm which can detect and track
multiple objects, and update target model automatically.The
contributions of this paper as follow: Firstly,we use color
histogram(HC) and histogram of orientated gradients(HOG)
to represent the objects,model update is realized under the
frame of kalman filter and gaussian model; secondly we use
Gaussian Mixture Model(GMM) and Bhattacharyya distance
to detect object appearance. Particle filter with combined
features and model update mechanism can improve tracking
effects.Experiments on video sequences demonstrate that mul-
tiple objects tracking based on improved algorithm have good
performance.
I. INTRODUCTION
Visual object tracking is an important task in the field
of computer vision. It is widely used in motion-based
recognition,automated surveillance, human-computer inter-
action,vehicle navigation and traffic monitoring [1], [2].
Color is one of the most widely used appropriate image
features in object tracking as its effectiveness and efficiency
[3], [4], [5]. Color histogram is robust to scale varia-
tion,rotation and small changes of forms.So color histogram
can be used in many tracking situations,as partial occlu-
sion and the object non-rigidity.However color histogram is
subject to similar color of its background and illumination
changes.So other characters,such as texture and contour
are employed to represent the object [6], [7], [8]. Spatial
distribution of color histogram is also used to represent the
object [9]. Recently,histogram of oriented gradients(HOG)
is employed to track object [10], [11], [12]. In [10] the
authors studied the issue of feature sets for human detection,
showing that locally normalized HOG descriptors provided
excellent performance relative to other existing feature sets
including wavelets.HOG represents the edges and local con-
tours so it is not subject to the color variation.We integrate
color histogram and histogram of oriented gradients(CHOG)
to achieve mutual complementarities of the two types of
feature.However CHOG can be applied to deal with some
appearance and shape changes,model update should be taken
into consideration.Over model update makes the tracker
sensitive to noise and occlusion,while under model update
makes the tracker lose the opportunities to follow objects
This work is supported by National Natural Science Foundation of
China(60736024),and Science Foundation of Ministry Education of China
under Grant No.1100310,No.1100070.
Yun Ji ZHAO is with College of Automation Science and Engi-
neering, South China University of Technology,Guangzhou 510640,China.
auzhaoyunji@163.com
Hai Long PEI is with College of Automation Science and Engi-
neering, South China University of Technology,Guangzhou 510640,China.
auhlpei@scut.edu.cn
appearance and shape changes.As in [13] we also use kalman
filter to do model update.
Deterministic methods localize the tracked object in each
frame by iteratively searching for a region which maximizes
the similarity measure between this region and target win-
dow.Comaniciu et al.[3], [14] employed mean shift algo-
rithm for object tracking.Mean shift algorithm is computa-
tionally efficient.However algorithms based on mean shift
may converges to a local maximum,they are also sensitive
to background distracts,occlusions,and objects quick mov-
ing.Among various of stochastic methods,Particle filter [11],
[12], [15], [16], [17] is widely used and makes the track-
ing system robustly.In [12] the author introduced switching
probabilistic principal component analysis model to update
the templates while in [18] color histogram was replace by
the sum of color histogram of target and estimate’s.Some of
color particle filter introduce model update as [18], while this
type of methods are subject to tracking drift.We introduce
kalman filter to do model update.In [19] the author proposed
multiple objects detection and tracking, object detection was
based on Bayesian estimation.Object detection based on
Bayesian estimation is delayed by a few frames. In this
article we introduced GMM and Bhattacharyya distance to
do objects appearance detection.
II. O
BJECT FEATURES AND MODEL UPDATE
A. Object features
Object features described by single cue do not work well
all the time.Many researches have employed multiple cues
for tracking system [11], [12], [15], [16], [17].In this article
we use HSV color histogram,each of them is composed of
𝑁
ℎ
× 𝑁
𝑠
+ 𝑁
𝑣
=8× 8+8 bins.
HOG [10] can describe the shape information of objects
which is calculated in gray-scale space as [11]. As SIFT
[20] we adopt the method to obtain the HOG,whilst we
introduce the weight function.To increase the reliability of
the HOG when boundary pixels belong to the background
smaller weights are assigned to the pixels further away from
the center.The weighting function as:
𝑤 = exp(−(𝑖 ∗ 𝑖 + 𝑗 ∗ 𝑗)
/
2 ∗ 𝛿
2
) (1)
where 𝑖 and 𝑗 is coordinate of pixels depending on the origin
which is the center of the computational domain.𝛿 is similar
to the one which is used to smooth the image.According to
the weight the HOG of the region will be obtained as follow:
𝐵
𝑢
𝑦
=
𝐼
∑
𝑖=1
𝑤(𝑥
𝑖
)𝛿(𝑏(𝑥
𝑖
) − 𝑢) (2)
2011 International Symposium on Intelligence Information Processing and Trusted Computing
978-0-7695-4498-4/11 $25.00 © 2011 IEEE
DOI 10.1109/IPTC.2011.10
9