Efficient Air-to-Air Drone Detection with Composite Multi-Dimensional Attention

Visual UAV detection has become a key technology in areas such as formation flight, low-altitude obstacle avoidance and anti-drone operations due to its affordablility, compact size and lightweight design. Air-to-air drone detection involves more complex background, unstable motion of source and target drones, small object sizes, varied shapes, substantial intensity variation, and occlusion, making it quite challenging. The visual attention mechanism shows promise in effectively addressing many of the aforementioned challenges. While some studies have incorporated attention algorithms into drone detection systems, there remains no systematic discussion of drone detection with multiple attention mechanisms. We explore the integration of attention mechanisms across three dimensions-scale attention, spatial attention, and task attention-into drone detection. Through detailed analysis, we assess their respective contributions and propose a novel visual attention drone detector. Experimental validation is performed on NPS-Drones and DUT-Anti-UAV datasets. The results show that the proposed drone detection algorithm based on attention mechanism exhibits significant advantages in both accuracy and processing speed.