A Discriminative Feature-Based Mean-shift Algorithm for Object Tracking

The mean-shift algorithm has been proved to be efficient for object tracking. Traditional mean-shift algorithm uses global color histogram features, regardless the features belong to the object or to the background, which will cause localization drift. In this paper, we propose a new algorithm which can overcome this disadvantage. Our hypothesis is that the features that best discriminate between object and background are also the best for tracking, and our tracking is based on these discriminative features. Features are chosen by separating the object from the background, using a voting strategy. Experimental results show that the proposed algorithm in this paper is more robust than the traditional mean-shift algorithm.