A Multi-features Based Particle Filtering Algorithm for Robust and Efficient Object Tracking - Multimedia and Signal Processing - page 8

A Multi-features Based Particle Filtering Algorithm
for Robust and Efficient Object Tracking

Shuang Ye 1,3 , Yanguo Zhao 1 , Feng Zheng 1 , and Zhan Song 1,2

1 Shenzhen Institutes of Advanced Technology,
Chinese Academy of Sciences, China
2 The Chinese University of Hong Kong, Hong Kong, China
3 Wuhan University of Technology, Wuhan, China
{shuang.ye,yg.zhao,feng.zheng,zhan.song}@siat.ac.cn

Abstract. This works presents a novel approach for robust and efficient object
tracking. To make the feature representation more robust, color and the local
binary pattern features are fused via a proposed scheme. The partial filter is
used for the feature tracking. To improve its efficiency, a mean shift based
method is introduced to decrease the required partials so as to decrease the
computation cost. With the robust multi-features description and boosted partial
filter algorithm, satisfied tracking results can be obtained via the experiments
with different datasets, and showed distinct improvements in both tracking
robustness and efficiency.

Keywords: object tracking, multi-features, mean-shift, particle filter.

1 Introduction

Object tracking has been an important research issue in computer vision domain.
Current tracking algorithms can be generally divided into two categories according to
Ref. 1: the certainty method and the random method. The former tracks the target by
looking for the optimal matching target, such as the mean-shift (MS) algorithm
[2] .The later tracks the target via state estimation, such as the particle filter (PF)
algorithm [3]
Various image features have been used for the target representation and tracking.
However, the tracking via sole feature is usually lack of robustness subject the
complicate background and scenarios. Much of recent methods have focused on the
fusion of multiple image features. [4-12] In Ref. 4, a PF-based method combined
with multiple image cues is proposed for the object tracking, but its efficiency still
needs improvement for real-time application. In Ref. 7, a geometric particle filter
algorithm is presented based on the affine group with optimal importance functions.
In Ref. 12, a new approach combined with MS of regional color distribution and the
PF algorithms is introduced for the efficient object tracking.
In this paper, a multi-features based approach for the efficient and robust object
tracking is presented. The color and local binary pattern (LBP) features are adopted