Visual Tracking with Dynamic Model Update and Results Fusion

Yu Zhu, Jing Wen, Liang Zhang, Yi Wang

Abstract: Sometimes the result of single tracker can be unreliable under some situation like illumination variation, occlusion, object size change, etc. Combining multiple estimates is a usually strategy to improve the performance of visual tracking, the ensemble approach can combine the advantages of difference models and overcome this limitation. In order to better fuse the results, we propose an adaptively fusion method that can select the weight of each track result automatically. Moreover, we propose an adaptively update strategy to avoid the “drift” during tracking. The assemble method and update strategy are selected by analyzing the situation of tracking response map. We expand Staple using our methods and evaluate the performance on famous object tracking benchmark. Experimental results show that our proposed method outperforms state-of-the-art tracking methods.