Learning Smooth Target-Aware Spatially Regularized Correlation Filters for UAV Tracking

Spatially regularized correlation filters (SRCF) have recently received increasing interest for Unmanned Aerial Vehicle (UAV) tracking due to their promising results. While the choice of spatial weight matrices is vital for the success of SRCF methods, they are generally learned only with the training samples in current frame, which results in time-discontinuous spatial weight matrices in neighboring frames, thus degrading the CF models. In this paper, we propose a Smooth Target-Aware Spatially Regularized Correlation Filter (STASRCF) framework for UAV tracking. Specifically, we first obtain the initial target-Aware spatial weight matrix in each frame by employing the image segmentation techniques for separating the target from the background, then multiple adaptive spatial regularization terms are integrated into the CF framework for jointly updating the spatial weight matrices and CF models. In this way, time-continuous spatial weight matrices and robust CFs can be learned during tracking, thereby benefiting the tracking performance. In addition, we suggest an Alternating Direction Method of Multipliers (ADMM) method for solving STASRCF efficiently, in which each sub-problem has a closed-form solution. Experiments on multiple UAV datasets show that STASRCF can not only surpass the baseline CSR-DCF by an average AUC gain of 1.9%, but also perform favorably against other state-of-The-Art CF trackers.