Sei sulla pagina 1di 4

2008 International Conference on Computer Science and Software Engineering

Multiple Non-Rigid Objects Tracking by Modified Kernel Particle Filter

CHENG Xiao-ping, CHEN Jia-shu


Department of Computer Science and Information Technology, SouthWest University,
Chongqing, China 400715
xpcheng@swu.edu.cn,
swucjs@swu.edu.cn

Abstract a more effective allocation of particles. The gradient


estimation and particle allocation is implemented by
This paper proposed a new method for multi-target the mean shift algorithm [1], [11]. [2] Which improved
tracking in video sequences by combining two trackers, sampling efficiency and the visual tracking robustness.
sum-of-squared differences (SSD) and kernel particle Though Kernel Particle Filter performs well on
filter (KPF). In our work, the idea of Object Likelihood sequences with relatively small object displacement, its
Value of pixel is proposed. Instead of using direct performance is not guaranteed when the objects move
propagation resample result from the previous sample fast as well as when they undergo partial occlusion,
set, a weighted SSD displacement is used for especially during long and continuous full occlusion.
reinitializing and resample before next KPF iteration. The motion pattern of objects may changes
Experiment results on soccer athletes tracking show dramatically in the situation listed above. Here, we
that the combination of SSD with KPF tracker forms a attempt to improve the performance of KPF tracker
simple and powerful multiple non-rigid object tracking when the object undergoes large displacements (when
system. the object regions do not overlap between the
consecutive frames) and in the event of partial /full
1. Introduction occlusion. The problem of large displacements is
tackled by cascading an SSD (sum-of-squared
Object tracking is an important task in many differences) tracker with the KPF tracker.
computer vision applications including surveillance,
gesture recognition, smart rooms, vehicle tracking, The outline of this article is as follows. In Section2
augmented reality, video compression, and medical we briefly give the key steps of the Kernel Particle
imaging, etc. The tracking of real-world objects is a Filter and in Section 3 we describe our modification of
challenging problem due to the presence of noise, SSD algorithm. In section 4, the SSD-KPF tracking
occlusion, clutter and dynamic changes in the scene system we proposed is discussed in detail and in
other than the motion of objects of interest. A variety Section 5 we present some experimental results.
of tracking algorithms have been proposed and to Finally, in Section 6, we summarize our research.
overcome these difficulties. The particle filters, also
known as condensation or sequential Monte Carlo, 2. Kernel Particle Filter
have proved to be powerful tools for object tracking
[7]. The current density of the state is represented by a 2.1 Estimation of Posterior PDF and the
set of random samples with associated weights and the gradient of PDF
new density is computed based on these samples and
weights. Kernel Particle Filter (KPF) was proposed Kernel density estimate (KDE) [10] is used to form
because the standard particle filter often fails to a continuous estimate of the posterior in order to
produce a particle set that captures the “irregular” facilitate gradient estimation. The target is
motion under some dynamic system conditions [2]. In characterized through the KDE. Given a particle set at
KPF, a kernel density estimate (KDE) [10] is used to n N
time t : st ={ st }n =1 and associated weights {πt
n N
approximate the posterior PDF. In particular, KPF }
n=1 ,
estimates the gradient of the kernel density and moves the KDE approximates the posterior PDF with kernel
particles toward the modes of the posterior, leading to K can be formulated as

978-0-7695-3336-0/08 $25.00 © 2008 IEEE 903


DOI 10.1109/CSSE.2008.1173
N ( n) approximated density modes, but the weight will
pˆٛ( Χt |Ζ1:t ) = ∑ K h ( Χt − st )π tn (1) contain a factor that offsets this effect.
n =1
Where K h is the kernel scaled by the width h ;
3 Modified SSD tracker
⎛ ( n) ⎞
( n) 1 ⎜ Χt − st ⎟
K h ( Χt − st ) = K (2)
h nx ⎜⎝ h ⎟

SSD tracker localizes the object in a given search
window of the successive frame based on minimum
The kernel and width h are chosen so as to minimize distance criterion between the target and candidate
the Mean Integrated Square Error (MISE) between the ˆ
posterior PDF and the corresponding kernel estimate. object images. Given the state estimate Χ t −1 at
Given the approximation of the posterior PDF, the time t − 1 , the SSD-based displacement estimate is
steepest direction toward the mode of the posterior ssd
could be searched, which can be found employing the D = arg m in ∑ [ I ( i+ Χ
ˆٛ ˆٛ
t − 1 + j ) − I t − 1 ( Χ t − 1 + j )]
2
t j∈ R
t ( 7)
mean shift procedure [11]. In this procedure, each i∈ S

particle is moved to its sample mean determined by Where I t and I t −1 are the two consecutive intensity
N (l ) l ( n ) (l )
∑ st π t g ( st − st ) images, S is the search window, and R is the ellipse
( n ) l =1 modeled the object.
st′ = (3)
N l ( n ) (l )
∑ π t g ( st − st )
l =1 3.1 Object Likelihood Factor and Weighted
After several iterations, the vector s t′ ( n ) − s t( n ) Displacement Estimation
would be in the gradient direction of (1) [11].
In the initialize step, there must be some method
2.2 Particle Reweighting (either manually or automatically) to segment object to
tracking from the background. Since the object is
(roughly) segmented in this step, it is relatively easy to
When particles change their positions, the new obtain color distribution of the object, which generally
particles do not follow the posterior distribution significantly different from that of the background. We
anymore. This is circumvented by re-weighting the can employ this characteristic to determine whether a
particles. Denote the particle set after the i th mean pixel belong to the tracking object. We define Object
shift procedure at time t as {st(,⋅i) } . After each mean Likelihood Factor (OLF) of a pixel in tracking scene at
time t as
shift procedure, the weight is recomputed as the
posterior density evaluated at the new particle
m ax { f i , o , ε } (7)
L ( i ) = lo g
positions augmented with a particle density balancing
t m ax { f i , b ,ε }
factor [2, 3]
(n) Where f i ,o is the probability that a pixel belong to the
p ( st ,i | Ζ1:t )
π t(,ni ) = (4) object, while f i ,b is the probability that a pixel belong
(n)
qt ,i ( st ,i )
to background. f i ,(⋅) is r-g-b combinational PDF, ε
Where the denominator is the new proposal density is a non-zero small positive number. It is evidently
that captures the non-uniformity of the new particle
set. if fi ,o > fi ,b , Lt (i ) is positive , else it is negative.
N Our view is that during computation of SSD
(l )
qt ,i ( Χt )= ∑ K h ( Χt − st ,i ) (5) displacement, the contribution of pixels in matching
l =1 template should not be taken as equal. The pixel more
Thus the posterior density can be given by likely on the object should assign greater weights.
(n) (n) N (n) (l ) Thus we have the following Weighted Displacement
p( st |Ζ1:t )∝ p( Ζt |st ) ∑ Kh ( st −st −1)πtn−1 (6)
l =1
Estimation
When mean shift is performed on st , the next
generation of particles will concentrate more on the

904
ssd ˆٛ + j ) − I ˆٛ 2 2 Compute combinational PDF of the object and
D = arg min ∑ Wt [ It ( D +Χ t −1 t −1 ( Χ t −1 + j )] (9) (9)
t
D∈S j∈R background,
3 Compute pixel OLF value.
In which 4 Compute a weighted SSD displacement Dt
ssd
by (11)
⎧ L (i ) Lt (i )>θ 5 Update particle set st′,0
( ⋅)
= st(,0⋅) + D(10)
ssd
Wt (i ) =⎨ t (10) t
⎩ 0 else
6 Do Mean-Shift KPF tracking procedure using those
θ is the threshold of the likelihood. In our system, particles with Li(t)> threshold,
7 Go step2 till the last frame processed.
θ =0.8. With the information of Lt (i ) , we need not
compute every pixel of template. In the case of partial 5 Experiments
occlusion, the occluded particles with zero OLF value
would be discarded and only those particles really The proposed algorithm was applied to many
contribute to tracking process left. sequences and found it work very robustly despite the
difficulties of suddenly change of velocity, background
3.2 Occlusion Order Determination clutter, and temporary occlusion.
The first experiment shows our approach is robust to
To determine the occlusion order is important to resolve the motion correspondence even under the
interpret the interaction among objects during background contained similar targets when tracking a
occlusion. The Dtssd originated from SSD represents single object (see Fig. 2). The tracker can still keep
pixel-wise intensity conservation of a search blob. tracking the soccer player on the right after walking
Figure 1 shows two objects overlapped each other. through a crowd of soccer players with similar clothing
Object A is above on B. We can obtain more color.
appearance information of A than that of B obviously.
In other words, object A conserved more intensity than
B. It is easy to assume that Dtssd − A > Dtssd − B is true in
generic situation. Inversely, we can follow occlusion
order from the size relation of the objects during
occlusion. To improve the confidence of the induced
ssd
result, we can set a threshold value DT . If

Dtssd − A ≥ Dtssd − B + DTssd (11)


is true, we believe object A is above on B.
Figure2. Tracking result of a soccer player passing-by a group of
soccer players. Frames shown 8, 10, 19, 21

The second example (in Fig. 3) shows the result of


multi-target tracking. There are several players going
through the scene in different directions, which result
many instances of their passing-by each other. In order
to test whether our algorithm can track objects
undergoes large displacement, we set a low FPS rate at
3fps. Our algorithm performs well as expected.

Figure1. A scenario of two targets overlapped.

4 SSD Modified KPF Tracker

The idea discussed above implemented as follows:


1 Initialized searching window according to object size

905
With help from SSD, the combined tracker can
address situations when objects undergo long and
continuous occlusion. It is expected the successfully
“fusion” of different type of tracking method will
further improve the performance of tracking system.

[1] D.Comaniciu,V. Ramesh, and P. Meer. Kernel-based


object tracking. In IEEE Trans. Pattern Anal. Machine
Intell.25 (5) (2003) 564-577.
[2] C.Chang and R. Ansari. Kernel particle filter for visual
Fig .3 Tracking result of multiple soccer players. Frames shown 9, 11, tracking. In IEEE Signal processing Lett., 12(3):242-
13, 15, 18, 20 245, March 2005.
The third example (in Fig. 4) is a segment includes 2 [3] C.Chang and R. Ansari. Multiple objects tracking with
players interacting from a long sequence. Player 1 is kernel particle filter. In IEEE CVPR 20-25 June 2005
occluded by player 2 in several frames. We can see our Page(s):566-573 vol.1.
tracker can still distinguish the one from the other, [4] Babu, RV and Perez P. Robust tracking with motion
estimation and local Kernel-based color modeling.
ssd − player1 ssd − player 2 ssd Image and Vision Computing 25 (8): 1205-1216 AUG 1
since the condition Dt ≥D
t
+D
T
is
2007.
always true during occlusion. In frame 12, it can still [5] T. B. Moeslund and A. Hilton. A survey of advances in
track the player 1 after long and continuous almost full vision-based human motion capture and analysis.
occlusion. Comput. Vis. Image Understanding, vol.104, no.2-3, pp.
90-127, 2006.
[6] K. Okuma, and A. Taleghani. A boosted particle filter:
Multitarget Detection and Tracking. In ECCV, 2004.
[7] Pérez, P., Hue. C, Vermaak, J., Gangnet, M.: Color-
Based Probabilistic Tracking. In ECCV,2002.
[8] M. Isard and A. Blake. Condensation——conditional
density propagation for visual tracking. Int. J. Comput.
Vision, vol.29, no. 1, pp. 5-28, 1998.
[9] S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp.
A tutorial on particle filters for on-line non-linear/non-
Gaussian Bayesian tracking. In IEEE Trans. Signal
Process., Vol. 50, no. 2, pp. 174-189, 2002.
[10] B. W. Silverman. Density Estimation for Statistics and
Figure 4. Tracking result of two soccer players. The player in blue Data Analysis. London, U .K.: Chapman & Hall, 1986.
ellipses was occluded by the other player in several continuous [11] Y. Cheng. Mean shift, mode seeking, and clustering. In
frames. But he can still be captured after he split from the group. IEEE Trans. Pattern Anal. Mach. Intell., vol. 17, no. 8,
Frames shown 1, 4, 12 pp. 790-799, Aug. 1995.
[12] D. Comaniciu, V. Ramesh, and P. Meer. Real-time
6 Conclusions tracking of non-rigid objects using mean shift. In proc.
IEEE Conf. on Computer Vision and Pattern
Recognition, Hilton Head, SC, volumeⅡ, June 2000, pp.
Multiple non-rigid objects tracking such as athletes 142-149.
tracking is quite a challenge for computer vision. We [13] B. Han, D. Comaniciu, Y. Zhu, and L. Davis.
believe single type tracker always have its own Incremental density approximation and kernel-based
limitation. In this paper, we have proposed an efficient Bayesian filtering for object tracking. In IEEE CVPR,
visual tracker by coupling SSD and KPF, which have Jun. 2004.
complementary properties. The key idea is to improve [14] C. Musso, N. Oudjane, and F. LeGland. Improving
KPF particle set update by a weighted SSD algorithm, regularized particle filters. In Sequential Monte Carlo
which moves the particles in KPF tracker to the most Methods in practice, A. Doucet, J. F. G. de Freitas, and
dominant modes quickly, since SSD give direct object N. J. Gordan, Eds. New York: Springer-Verlag, 2001.
[15] D. Comaniciu and P. Meer. Mean shift: A robust
pixel motion information between frames. While KPF approach toward feature space analysis. In IEEE Trans.
trackers are used here for finer adjustments of object Pattern Anal. Machine Intell., vol. 24, no. 5, pp. 603-619,
location, 2002.

906

Potrebbero piacerti anche