Self-repairing object tracking method by adopting multi-level features

Visual object tracking is a fundamental problem in computer vision community and has been studied for decades. Trackers are prone to drift over time without other information. In this paper, we propose a self-repairing online object tracking algorithm based on different level of features. The fine-grained low-level features are used to locate the specific object in each frame and the coarse-grained high-level features are used to describe the category-level representation. We design a tracking kernel updating mechanism based on category-level description to revise the online tracking drift. We tested our proposed algorithm on OTB-50 dataset and compared the proposed method with some popular real-time online tracking algorithms. Experimental results demonstrated the effectiveness of our proposed method.