A novel approach for visual tracking based on occlusion recognition

Abstract

Visual object tracking is an extremely challenging task. Many existing trackers cannot handle various challenges simultaneously. In this paper, we propose a novel tracking framework based on an occlusion recognition mechanism to improve the performance in occlusion situations. Firstly, we design an occlusion recognition mechanism based on patch pool and local correlation to describe the occlusion of objects in each frame of an image sequence. Secondly, taking advantage of the occlusion recognition mechanism, we construct a specific training set to train the filter. Thirdly, combining global correlation, we implement our own tracker based on the traditional discriminative correlation filters. Finally, we evaluate it on both OTB and VOT platforms, and the experimental results demonstrate that our design is advanced and effective.