Part-based correlation filter tracking by exploiting the similarity and contribution of reliable parts

Abstract Challenges like occlusion, background clutter, illumination and scale variation make visual object tracking a tough problem in computer vision. Part-based tracking methods have been widely used to solve the partial occlusion issue, different from various holistic appearance model based trackers, it combines the local appearances of the target to build a robust global appearance. In this paper, we propose a novel object tracking method based on multi-part scheme and Bayesian framework, using the Kernelized Correlation Filter (KCF) as the base tracker. Additionally, two robust reliability metrics are proposed: the first one utilizes the similarity between the corresponding parts in consecutive frames to measure the reliability; while the second one measures the contribution that each part made for the global target object. Besides, adaptive updating strategy and scale estimation which employ the relationships among parts are proposed to deal with appearance modeling. Extensive experiments have been conducted on latest benchmark to demonstrate that our method outperforms state-of-the-art trackers.