A Timely Occlusion Detection Based on Mean Shift Algorithm

Mean shift algorithm has attracted much attention in computer vision and has recently shown promising performance in the challenging problem of visual tracking, but it is difficult to deal with occlusion. In this paper, a timely occlusion object detection based on mean shift is proposed. By analyzing occlusion process, it is evident to find that occluded size is increasing and occlusion patch lies to edge of objects at the beginning. So object model is divided into several parts. In order to reduce computation, only edge patches is considered. If Bhattacharyya coefficient of one patch decreases greatly and other patches change faintly, it means that object is occluded in this area. In this method, object model is divided into two parts, one is occlusion part, the other is no occlusion part, the no occlusion part of the object can be obtained to track object continuatively until object is occluded totally. Experiments show that compared with whole object model judges, it is timely to detect occlusion and deals with occlusion successfully.