Patch-based object tracking using the local robust histogram and background estimation

A novel patch-based algorithm for robust object tracking is proposed in this study. The patches of the appearance model are represented by the proposed local robust histogram. Then, the background model is constructed by a set of new spatial probability maps in a surrounding "context window". For a new testing frame, the vote maps that are obtained by matching the target patches independently are fused for determining the new location of the object. Then, a two-stage estimation method is proposed to estimate the probability of the pixels belonging to the target in the new location. The patches are classified into foreground patches and occluded patches. At last, a dynamic updating scheme is proposed to address appearance variations and alleviate tracking drift. Experiments and evaluations on various challenging image sequences are performed, and the results show that the proposed algorithm performs favorably against other state-of-the-art methods.