4.7 Article

ReCF: Exploiting Response Reasoning for Correlation Filters in Real-Time UAV Tracking

期刊

出版社

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/TITS.2021.3094654

关键词

Training; Target tracking; Unmanned aerial vehicles; Real-time systems; Correlation; Feature extraction; Cognition; Real-time UAV tracking; discriminative correlation filter; adaptive response reasoning

资金

  1. National Natural Science Foundation of China [61806148]
  2. Natural Science Foundation of Shanghai [20ZR1460100]

向作者/读者索取更多资源

An adaptive response reasoning approach is proposed for CF learning in UAV object tracking, which leverages temporal information in filter training and significantly enhances the robustness of the tracker. The method goes beyond standard response consistency requirements, creating an auxiliary label for the current sample and learning a generic relationship between previous and current filters to achieve self-regulated filter updating and improved discriminability.
Object tracking is a fundamental task for the visual perception system on the intelligent unmanned aerial vehicle (UAV). The high efficiency of correlation filter (CF) based trackers has advanced the widespread development of online UAV object tracking. This kind of method can effectively train a filter to discriminate the target from the background. However, most CF-based methods require a fixed label function over all the previous samples, leading to over-fitting and filter degradation, especially in complex drone scenarios. To address this problem, a novel adaptive response reasoning approach is proposed for CF learning. It can leverage temporal information in filter training and significantly promote the robustness of the tracker. Specifically, the proposed response reasoning method goes beyond the standard response consistency requirement and constructs an auxiliary label of the current sample. Besides, it helps learn a generic relationship between the previous and current filters, thereby realizing self-regulated filter updating and enhancing the discriminability of the filter. Extensive experiments on four well-known challenging UAV tracking benchmarks with 278 videos sequences show that the presented method yields superior results to 40 state-of-the-art trackers with real-time performance on a single CPU, which is suitable for UAV online tracking missions.

作者

我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。

评论

主要评分

4.7
评分不足

次要评分

新颖性
-
重要性
-
科学严谨性
-
评价这篇论文

推荐

暂无数据
暂无数据