Robust visual tracking using multiple cues and improved particle filter
Guodong Tian, Bo Yang, Hongling Wang · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2009
A robust visual tracking method which can be used in complex environments is presented in this paper. The color cue and the shape cue are utilized to represent the target and fused together by democratic integration method. The multi-cue object representation is incorporated into the framework of particle filter which is a powerful probabilistic method for visual tracking. To each sample of the particle filter a mean shift operation is applied, which make the samples more effective such that the number of particles needed is significantly decreased. Unlike regular mean shift, in our method the number of mean shift iterations is limited according to the reliability of the color cue for two purposes. One is to prevent the particles from being misled by mean shift when the color cue is unreliable. The other is to reduce the waste of computation. Experimental results show that our method greatly improves the robustness and reduces the computational cost compared with the state-of-art methods.