Learning Dynamic Distractor-Repressed Correlation Filter for Real-Time UAV Tracking

Abstract

With high-efficiency computing advantages and desirable tracking accuracy, discriminative correlation filters (DCFs) have been widely utilized in UAV tracking, leading to substantial progress. However, in some intricate scenarios (e.g., similar objects or backgrounds, background clutter), DCF-based trackers are prone to generating low-reliability response maps influenced by surrounding response distractors, thereby reducing tracking robustness. Furthermore, the limited computational resources and endurance on UAV platforms drive DCF-based trackers to exhibit real-time and reliable tracking performance. To address the aforementioned issues, a dynamic distractor-repressed correlation filter (DDRCF) is proposed. First, a dynamic distractor-repressed regularization is introduced into the DCF framework. Then, a new objective function is formulated to tune the penalty intensity of the distractor-repressed regularization module. Furthermore, a novel response map variation evaluation mechanism is used to dynamically tune the distractor-repressed regularization coefficient to adapt to omnipresent appearance variations. Considerable and exhaustive experiments on four prevailing UAV benchmarks, i.e., UAV123@10fps, UAVTrack112, DTB70 and UAVDT, validate that the proposed DDRCF tracker is superior to other state-of-the-art trackers. Moreover, the proposed method can achieve a tracking speed of 59 FPS on a CPU, meeting the requirements of real-time aerial tracking.