Stereo-based tracking of multiple overlapping persons

This paper describes a method of tracking multiple persons with occlusions using stereo. We previously developed an accurate and stable tracking method using overlapping silhouette templates which considers how persons overlap in the image. It realized a fast tracking by using an approximated likelihood map based on kernel density estimation. The method, however, treated only two overlapping persons. In this paper, we propose an improvement of approximation applicable to the case where three or more persons overlap. Experimental results show that the proposed method can track persons stably even when three persons overlap in the image.