A Reinforced Deep Reinforcement Learning Method for UAV Target Tracking

Abstract

Abstract In response to challenges faced during unmanned aerial vehicle (UAV) image target tracking, such as target orientation changes, occlusion variations, and limited sample diversity, we propose a reinforced deep reinforcement network-based algorithm for UAV aerial image target tracking. Addressing the issue of limited sample diversity, we enhance the original samples through occlusion and rotation. Next, we construct an improved deep belief network to distill representative deep features, ensuring the deep model can accurately and rapidly adapt to target shape changes and obtain the pre-localization region of the target to be detected. We then employ a reinforcement learning algorithm for precise target localization. Finally, a deep forest classifier is utilized to output the final target tracking results. Comparative experiments on multiple datasets demonstrate that the proposed algorithm achieves high tracking accuracy, is capable of handling target rotation and occlusion, and exhibits excellent precision and robustness.