A new similarity measure and back-projection scheme for robust object tracking

We propose a new object similarity measure, which can boost the performance of the mean-shift based algorithms for robust object tracking. The proposed scheme can better discriminate between different objects, compared to the commonly used measure based on Bhattacharyya coefficient optimization. The improvement is particularly noticeable when the probability distribution function of the tracked object covers a wider range of the chosen feature space. Based on this new similarity measure, we present a back-projection scheme to create probability images in which the object of interest stands out clearly, and can be tracked robustly using the mean-shift algorithm. The results are remarkably better than the traditional mean-shift tracking, especially when the object moves fast and there is very little overlap between the object positions in successive frames.