Distractor-Aware Long-Term Correlation Tracking Based on Information Entropy Weighted Feature

Correlation filter based tracking algorithms have received significant attention in visual tracking due to their accuracy and speed. However, it is susceptible to distractors and performs poor when it is faced with occlusion and severe deformation. In this paper, we introduce a distractor-aware long-term correlation tracking to distinguish the target and the background. Firstly, the information entropy weighted feature is utilized to deal with different scenarios dynamically. Secondly, in order to facilitate the tracker to distinguish the target from the distractors which lead to inaccurate localization, we propose a novel distractor-aware mask processing. We sample distractor patches and incorporate a binary mask to attenuate the possible negative influences. Thirdly, an improved fast scale searching scheme based on dichotomy is incorporated to provide a tradeoff between the accuracy and efficiency. Finally, the localization ambiguity is detected by using a new criterion based on the variance. It is concluded that introducing the re-detection technique and solving the occlusion play a critical role to facilitate the tracking progress. Furthermore, an adaptive model update strategy is utilized to update the long-term model for alleviating the model drift. Experiments on benchmark database demonstrate that the proposed method achieves good performance compared with other trackers and the strategy can be applied to other trackers.