Distracter-aware correlation filter tracking

Abstract

We propose a distracter-aware Correlation Filters (CF) tracking algorithm, which exploits the information of dis-tracters to enhance the robustness of the tracker. Although most existing correlation filters based trackers achieve accurate tracking results, they may be less effective when similar distracters appear in the background. To this end, the proposed algorithm not only take the information of the target into consideration but also pay attention to the information of the distracters in the background. We first detect the distracters based on the response of CF model and then design a label map based on the information of the detected distracters. Unlike most existing CF based trackers which direct use a Gaussian shape label map, the proposed algorithm design a distracter-aware label map which makes the trained CF model effective to handle distracters. The proposed algorithm has several compelling advantages: it detects distracters, captures the discriminative information, which is crucial for robust tracking, enhances the robustness. We evaluate our proposed algorithm on the public OTB datasets, which including 50 sequences, and compare it with several state-of-the-art trackers. The comparable results show the effectiveness of the proposed algorithm.