Learning target-aware background-suppressed correlation filters with dual regression for real-time UAV tracking

Abstract

Traditional discriminative correlation filters have been rapidly developed in UAV tracking field due to their effectiveness and computational efficiency. Unfortunately, these methods are limited by unwanted boundary effects, which decrease the discriminative power between the target and background. Besides, due to the fast target appearance change and complex scenarios, such as background clutter and similar object, there often exists distractors around the response peak, which may cause tracking drift or even failure. In this paper, we propose a novel correlation filter method with dual regression, which aims at obtaining high-quality and reliable response maps for more robust tracking. Specifically, benefiting from a saliency detection method, the target feature is efficiently acquired from the global feature. Then, employing the dual regression strategy, these features are used to regress dual filters, that is, the target filter and the global filter. Response maps generated by dual filters, i.e., the target and global response maps, are merged in the detection phase to increase the response value of the target. Furthermore, a novel response-level background suppression regularization is proposed to solve boundary effects. Through the mutual restriction between the target and global response maps, the discrimination of dual filters can be promoted. We perform extensive and comprehensive analysis on three challenging UAV tracking benchmarks. Results confirm that the dual regression strategy and the background suppression regularization can facilitate the tracking accuracy improvement. The proposed tracker has comparable performance against other 27 state-of-the-art trackers while running at ∼34 FPS on a single CPU.