Behavior cloning for active object detection in UAV remote sensing

Active object detection (AOD) in Unmanned Aerial Vehicle (UAV) remote sensing faces challenges of multi-scale targets and partial occlusions due to varying viewpoints and scales. This paper presents a novel approach that integrates behavior cloning with AOD to enhance detection performance in complex aerial environments. By leveraging expert demonstrations generated through the A-star algorithm, our method learns optimal camera control strategies without the complexity of traditional reinforcement learning. The proposed framework combines shallow feature pyramid network features with detector outputs and historical actions to make decisions about camera positioning. Experimental results on two datasets demonstrate that our approach achieves competitive mean Average Precision scores compared to existing methods while significantly improving processing efficiency, with up to 63% increase in frames per second. This work represents a practical solution for real-time UAV object detection applications where both accuracy and computational efficiency are crucial.