Robust Tracking via Dual Constrained Filters

Abstract

In this paper, we propose a novel correlation filter framework constrained by dual filters. The Minimum Output Sum of Squared Error (MOSSE) filter is the unbiased estimate of the filter which easily to cause overfitting. The trained filter by linear ridge regression is the biased estimate of the filter which can deal with the overfitting. We combine the advantages of the two filters to constrain the trained filter which optimizes our model. To deal with background clutter, clipping background patches around the target position up, down, left, and right, we add the cropped background patches to the learning filter. To overcome the challenge of occlusion, we introduce a novel criterion, Average Peak-to-Correlation Energy (APCE). Extensive experiments on the CVPR 2013 Benchmark well demonstrate that our tracker can effectively solve the background clutter and occlusion. Both quantitative analysis and qualitative analysis show that our tracker outperforms some state-of-the-art trackers.