Adaptive Updating Siamese Network with Like-Hood Estimation for Surveillance Video Object Tracking
Zhenxian Zheng, Yang Yi, Jin-Long Shen, Jiahao Zhang · 2019
Surveillance video object tracking is considered vital with the expanding of surveillance services. Many recent trackers have achieved advanced performance while model drift is still a challenging topic in surveillance scenes because of background interference and object appearance variation. In this paper, we propose an ensemble framework consisting of two branches to avoid model drift. In the correlation branch, a well-designed Siamese network is presented which employs a spatial channel attention module to promote tracking performance and adopts novel criterions to adap tively update filter templates thus preventing model drift. In the like-hood branch, a like-hood estimation method is proposed to correct object position which effectively reduces similar backgrounds' interference. Then two branches combine to get the ultimate response. Experimental results on four challenging surveillance video related benchmarks show that our approach has state-of-the-art tracking performance while runs at 65 frames per second speeds showing great potential in practical tracking application.