SiamTADT: A Task-Aware Drone Tracker for Aerial Autonomous Vehicles

State-of-the-art (SOTA) drone trackers struggle with small target bounding box regression and target classification amidst background interference in drone tracking scenarios. These challenges arise from a “task-unaware” conventional tracking methodology that treats bounding box regression and classification as indistinct tasks, neglecting their differing feature requirements. This paper introduces a “<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><b>task-aware</b></i>” tracking methodology, addressing these issues by discriminating between the two subtasks and assigning them their optimal feature representations. The methodology includes a “<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><b>task-aware</b></i>” encoder, a “<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><b>task-aware</b></i>” decoder, and a mutual-self training strategy. The encoder provides tailored feature representations for each subtask, while the decoder employs distinct structures to leverage these task-specific features. The mutual-self training strategy further enhances performance by fostering synergy between the subtasks. Experimental results demonstrate the effectiveness and robustness of the proposed approach. It outperforms the SOTA drone tracker by 2.3% and 3.8% across six drone tracking datasets and surpasses its baseline by 12% and 9% on average. Additionally, tests on conventional tracking datasets show the method's robustness, matching SOTA conventional trackers while improving the baseline by 28% and 13.5% across two metrics. This innovative method bridges the gap between task-specific needs and feature usage, advancing drone tracking performance. Codes and results will be publicly available.