Robust and efficient fragments-based tracking using mean shift

Representing a target with multiple image fragments has been used for tracking in a video sequence and proved to be able to preserve the spatial information. However, this kind of target description poses additional challenges in terms of efficiency of a video target tracking. In this paper, we present a novel algorithm that addresses this issue. We propose a fusion scheme to fuse multiple spatially distributed fragments effectively. Under the fusion scheme, we derive a mean shift type algorithm which allows efficient target tracking with very low computational overhead. In order to properly fragment the human target, we propose a GrabCut-based method to divide the target into multiple distinctive fragments. The proposed algorithm is evaluated on extensive challenging sequences. The experimental results demonstrate that it is effective and efficient.