Short Communication
Particle filter based on Particle Swarm Optimization resampling for vision tracking
Jing Zhao
*
, Zhiyuan Li
College of Computer Science and Technology, Harbin Engineering University, Harbin 150001, China
article info
Keywords:
Object tracking
Particle filter
Particle Swarm Optimization
abstract
Particle filter is a powerful tool for vision tracking based on Sequential Monte Carlo framework. The core
of particle filter in vision tracking is how to allocate particles to a high posterior area. Particle Swarm
Optimization (PSO) is applied to find high likelihood area in this paper. PSO algorithm can search the
sample area around the last time object position depending on current observation. So, it can distribute
the particles in high likelihood area even though the dynamic model of the object cannot be obtained. Our
algorithm does not distribute the particles based on the weight of the particles last time like the sam-
pling-importance resampling (SIR). SIR usually allures particles distributed in wrong likelihood area par-
ticularly tracking in cluttered scene. Since that some particles have larger weight maybe illusive. We first
find the sample area by PSO algorithm, then we distribute the particles based on two different base points
in order to achieve diversity and convergence. Experimental results in several real-tracking scenarios
demonstrate that our algorithm surpasses the standard particle filter on both robustness and accuracy.
Ó 2010 Elsevier Ltd. All rights reserved.
1. Introduction
In recent years, there has been a great deal of interest in apply-
ing particle filtering by Arulampalam, Maskell, Gordon, and Clapp
(2002), also known as Condensation (Isard & Blake, 1998), to com-
puter vision problems. Applications on rigid or non-rigid object
tracking has demonstrated its usefulness (Choo & Fleet, 2001;
Zhang & Pece, 2003).
The most appealing aspect of PF is to maintain multiple hypoth-
esis of a state, which makes them more competent for heavily clut-
tered and complex scenes. However, PF are not always satisfactory,
especially when the irregular and abrupt motion renders a weak
dynamic model in real scenes (Chang & Ansari, 2005). There is a
motion model in the transition equation. The motion model can
help distribute the particles in the right sample area. But some-
times we do not know the motion model of the object. If we do
not use the motion model, the particles may be distributed in the
wrong sample area when the object is similar to background
(Zhang & Pece, 2003).
The resampling scheme in Condensation is sampling particles in
last time particle set according to their weight. The larger weight
the particle has, the more times the particle will be chosen to gen-
erate the next time particles. But that some particles have large
weight is illusive in clutter scene and when occlusion occurs. These
‘‘large” weight particles are far from the high likelihood area.
So they will abduct the particles distributed in the wrong area
piece by piece.
An improved strategy to overcome these problems is to design
better proposal distributions. An auxiliary particle filter (APF) is
one such example provided by Pitt and Shephard (1999), which
generates particles from an importance distribution depending
on a more recent observation. Its weakness is that it requires a
large number of particles. When the state transition density is
quite scattered and the likelihood varies significantly over the state
transition distribution, APF is not always effective.
Another enhanced tactic for PF has been extensively studied
since the pioneering work of Comaniciu, Ramesh, and Meer
(2003), who were the first to introduce mean-shift analysis to vi-
sual tracking. Following their work, Maggio and Cavallaro (2005),
Shan, Wei, Tan, and Ojardias (2004) subtly extended mean shift
to the particle-filter framework. The central idea of their algo-
rithms is to redistribute particles to their local mode of the poster-
ior density by mean-shift analysis, thereby possibly using fewer
particles to keep multiple hypotheses. The particles in this algo-
rithm have lack of diversity. It performs ineffectively when occlu-
sion occurs. Different from the work of Maggio and Cavallaro
(2005), Shan et al. (2004), Chang and Ansari (2005) presented an-
other method which approximated posterior density using kernel
density estimate (KDE) and then estimated the gradient of poster-
ior density by mean-shift analysis. Mean shift plays different roles
in the above two kinds of methods in that it was used for maximiz-
ing the similarity function between the target candidate and the
target model by Maggio and Cavallaro (2005), which was used as
mode seeking of posterior density in Chang and Ansari (2005). De-
0957-4174/$ - see front matter Ó 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.eswa.2010.05.086
* Corresponding author.
E-mail address: zhaoj@hrbeu.edu.cn (J. Zhao).
Expert Systems with Applications 37 (2010) 8910–8914
Contents lists available at ScienceDirect
Expert Systems with Applications
journal homepage: www.elsevier.com/locate/eswa