A Multiple Granular Cascaded Model of Object Tracking Under Surveillance Videos
Haoren Cui, Zhihua Wei, Pengyu Zhang, Di Zhang · 2018
Object tracking is one of the most important components in numerous applications of computer vision, and much progress has been made in recent years with efforts on sharing code and datasets. In this paper, we propose a multiple granular cascaded model of object tracking under surveillance video. For frames in surveillance video, we analyze the frames from different granularity levels. At the coarse level, we utilize Gaussian mixture model combined with a three-frame difference method to extract the foreground region of the frame. At the fine level, a cascaded multi-source feature fusion method is used to classify the region proposal and obtain the location of the target. Finally, we propose some optimization strategies for the online tracker updating. We conduct several comparative experiments on actual surveillance video. Experimental results demonstrate that the proposed model is able to obtain excellent results in practice.