EACOFT: An energy-aware correlation filter for visual tracking. (April 2021)

• Record Type: Journal Article
• Authors: Liu, Qiaoyuan; Ren, Jinchang; Wang, Yuru; Wu, Yuanbo; Sun, Haijiang; Zhao, Huimin
• Is Part Of: Pattern recognition. Volume 112(2021)
• Publication Date: 2021-04
• Subjects: Visual tracking -- Energy-aware correlation filter (EACOFT) -- Enhanced feature -- Top-down and bottom-up strategy; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107766
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Energy-aware-correlation-filter tracker to adaptively adjust the target for tracking. New strategy to reject low quality samples and ensure model discriminant ability. Combining bottom-up and top-down optimal strategy for training and robust tracking. Outperform many state-of-the-art trackers on several challenging datasets. Abstract: Correlation filter based trackers attribute to its calculation in the frequency domain can efficiently locate targets in a relatively fast speed. This characteristic however also limits its generalization in some specific scenarios. The reasons that they still fail to achieve superior performance to state-of-the-art (SOTA) trackers are possibly due to two main aspects. The first is that while tracking the objects whose energy is lower than the background, the tracker may occur drift or even lose the target. The second is that the biased samples may be inevitably selected for model training, which can easily lead to inaccurate tracking. To tackle these shortcomings, a novel energy-aware correlation filter (EACOFT) based tracking method is proposed, in our approach the energy between the foreground and the background is adaptively balanced, which enables the target of interest always having a higher energy than its background. The samples' qualities are also evaluated in real time, which ensures that the samples used for template training are always helpful with tracking. In addition, we also propose an optimal bottom-up and top-down combined strategy for template training, which plays an important role in improving both the effectiveness and robustness of tracking. As a result, our approach achieves a great improvement on the basis of the baseline tracker, especially under the background clutter and fast motion challenges. Extensive experiments over multiple tracking benchmarks demonstrate the superior performance of our proposed methodology in comparison to a number of the SOTA trackers.