Robust visual tracking via a hybrid correlation filter

In this paper, we propose a hybrid correlation filter based tracking method which depends on coupled interactions between a global filter and two local filters. Specifically, a local kernel feature with Gaussian curvature is developed to encode object appearance. Then the global filter and the two local filters independently track the target. The peak-to-sidelobe ratio (PSR) is employed to measure the reliability of the tracking results. Next, the global filter and the two local filters jointly determine the target position. In this way, the proposed hybrid model deals well with challenging situations, e.g., partial occlusion and scale changes. Experiments on large benchmark datasets show that our method performs favorably against state-of-the-art trackers.