Multi-object Particle Filter Tracking with Automatic Event
Analysis
Yifan Zhou
LaBRI, Université Bordeaux 1
351, cours de la Libération,
33405, Talence cedex, France
yifan.zhou@labri.fr
Jenny Benois-Pineau
LaBRI, Université Bordeaux 1
351, cours de la Libération,
33405, Talence cedex, France
jenny.benois@labri.fr
Henri Nicolas
LaBRI, Université Bordeaux 1
351, cours de la Libération,
33405, Talence cedex, France
henri.nicolas@labri.fr
ABSTRACT
The automatic video content analysis is an important step
to provide the content-based video coding, indexing and re-
trieval. It is also a key issue to the event analysis in video
surveillance. In this paper, an automatic event analysis ap-
proach is presented. It is based on our previous method
of Multi-object Particle Filter Tracking with Dual Consis-
tency Check. The multiple non-rigid objects are first tracked
individually in parallel by multi-resolution technique and
particle filter method. The events including object pres-
ence and occlusion identification are then detected and an-
alyzed by measuring the Goodness-of-Fit Coefficient based
on Schwartz’s inequality and the Backward Projection. The
method is then tested in different indoor and outdoor envi-
ronments with cluttered background. The experimental re-
sults show the robustness and the effectiveness of the method.
Categories and Subject Descriptors
I.4.9 [Computing Methodologies]: IMAGE PROCESS-
ING AND COMPUTER VISION—Applications
General Terms
Algorithms
Keywords
Particle filter, multiple non-rigid object tracking, event anal-
ysis
1. INTRODUCTION
With the development of object extraction techniques in
videos, there is an increasing interest in the automatic Video
Content Analysis (VCA)[4]. On one hand, it is an essential
step to provide content-based video coding, indexing and
retrieval. On the other hand, it is a key issue in automati-
cally efficient content analysis for video surveillance systems.
The video objects, which are directly related to the video
events, are the basic units used for analyzing the video con-
tent. Therefore, a fundamental task in the VCA is to extract
those events of interest from the videos [6].
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
ARTEMIS’10 October 29, 2010, Firenze, Italy.
Copyright 2010 ACM 978-1-4503-0163-3/10/10 ...$10.00.
There exists numerous works in detecting and identifying
the events in videos. A lot of them are based on the generic
visual properties of frames. For example, histogram change
[11] between two consecutive frames may indicate the oc-
currence of an event and the object trajectory [14] can be
used for the object motion analysis. Hence, the histogram
and the trajectory are the basic analytic elements in our
automatic event analysis method.
The event analysis is based on the object state estimate
obtained by our previous tracking method: Multi-resolution
Particle Filter Tracking with Dual Consistency Check (MOP-
FDCC) [18]. The inspiration of the tracking method came
from the work [15]. It is a Particle Filter (PF) method [3]
where the object appearance (histogram) is used to evalu-
ate the importance function of particles. We chose it for its
capacity to solve non-linear and non-Gaussian state-space
model, e.g., non-rigid object tracking. Unfortunately, this
tracking method is very sensible when tracking non-rigid ob-
jects in videos of low and variable frame-rate due to strong
motion magnitude of objects. Therefore, we integrated the
Multi-resolution technique to PF method to reduce this mag-
nitude [17]. The essential idea was to first quickly locate the
objects by the prediction at the smallest resolution level and
then refine it gradually at higher levels. Instead of track-
ing only on full resolution [15] or a certain optimal level
[5], our method involves all the levels. A Dual Consistency
Check (DCC) [18] was employed to control the sample de-
generacy problem [2]. Instead of using an effective sample
size [2] to evaluate the effectiveness of a set of particles,
the Kolmogrov-Smirnov (KS) test [12] was applied to check
the change of object appearance model. It then decides the
further processing step: updating object appearance, reini-
tializing object estimate or hiddenly tracking the object.
In this paper, we improve our previous work MRPFDCC
by integrating an automatic event analysis stage and we call
the overall method a “Multi-object Particle Filter Tracking
with Event Analysis” (MOPFEA). The event analysis stage
is composed of object presence and occlusion identification
step. It is realized by measuring the Goodness-of-Fit Coef-
ficient (GFC) based on Schwartz’s inequality and Backward
Projection. The method will be explained in details in Sec-
tion 2. The experimental results will be illustrated in Section
3 and the conclusion can be found at the last in Section 4.
2. TRACKING ALGORITHM
The principle of PF [3] for the object tracking relies on ap-
proximating the expectations of a state function of an object
by a weighted sum of particles. Suppose X
l,k
t
is the estimate
of the object k at the time t at the resolution l, computed by