Particle filter object tracking based on SIFT-Gabor Region Covariance Matrices

Abstract

Currently, object tracking is an important problem to computer vision community. It is usually performed in the context of higher-level applications aiming to accurately label and track target objects in frame sequences. However, video-based object tracking is very challenging, since the objects are easy to lose when illumination varies or occlusion occurs. To solve these problems, considering the SIFT and Gabor features perform robustly for objects representation, a novel method is proposed in which target model is constructed by SIFT-Gabor Region Covariance Matrices (SG-RCMs) and particle filter is used to track the object. In the tracking process, the target model is updated automatically according to the matching result between target model and candidate targets. Experimental results showed that the proposed approach tracks the object of which illumination and scale are drastically changing, effectively, accurately and robustly.