Learning residue-aware correlation filters and refining scale for real-time UAV tracking

Highlights：

• We propose a novel regularization to model the residue between two neighboring frames, resulting in what we call residue-aware correlation filters, which show better convergence properties in filter learning. Meanwhile, we add spatial and temporal regularizations to boot performance with little additional computational cost.

• We propose a novel scale estimation approach for DCF-based trackers by using the GrabCut algorithm to refine the discriminative scale estimates, which can be incorporated easily into any tracking method with discriminative scale estimation to improve precision and accuracy.

• We demonstrate the proposed methods on four UAV benchmarks, namely, UAV123@10fps, DTB70, UAVDT and Vistrone2018 (VisDrone2018-test-dev). Experimental results show that the proposed approaches achieves state-of-the-art performance.