Adaptive target response-based spatio-temporal regularized correlation filter for UAV-based object tracking

Unmanned Aerial Vehicles (UAVs)-based visual object tracking has drawn significant attention recently and is extensively used in military applications, aviation, security, and agriculture, to name a few. Discriminative Correlation Filter (DCF)-based trackers have gained significant recognition in UAV-based visual tracking due to their efficiency and tracking speed. However, the training samples generated by cyclic shifts in the correlation filter are not real samples. In many cases, these approximate samples differ significantly from actual samples. Due to this, the traditional correlation filter-based trackers have inherent boundary effects and filter degradation issues. The traditional assumption of a single-centred Gaussian target response may not be reliable in challenging situations such as fast motion or occlusion. Such unreliable training samples lead to tracker drift. This paper proposes a novel Adaptive Target Spatio-Temporal Regularized Correlation Filter (ATCF) tracker to rectify these issues. A simple yet effective energy function is developed by combining the adaptive spatio-temporal regularized correlation filter and the adaptive target response. The closed-form solution is obtained using the Alternating Direction Multiplier Method (ADMM). The target response and spatio-temporal regularization parameters are learned online. Besides, a novel detection-based re-tracking strategy is introduced to improve long-term tracking performance. The proposed tracker’s performance on four benchmark datasets, i.e. DTB70, UAV123@10fps, UAV20L, and UAVDT benchmarks, has proved its superiority over various state-of-the-art trackers in terms of accuracy and robustness while running in real time in a CPU environment.