A Robust Joint Space Tracking Method Based on Particle Filter

We propose a novel object tracking algorithm based on modeling the target appearance in a joint space. In contrast with traditional histogram-based trackers which discard all spatial information, the joint space takes both the photometric and spatial information into account. Within this joint space, the target is modeled in a Gaussian mixtures manner where a richer description of the target is captured to increase robustness in tracking. The particle filter procedure is employed for sampling and dealing with the problem of getting trapped in local optimum. Finally, the updating step enables tracking to continue under severe occlusion. Experimental results on a number of challenging sequences demonstrate the effectiveness of the technique.