A2MGA-track: lightweight motion-geometric association for real-time air-to-air multi-UAV tracking
Abstract
Air-to-air tracking of multiple unmanned aerial vehicles (UAVs) is an important visual perception task for low-altitude airspace supervision and aerial security. However, small target scales, weak appearance, complex relative motion, and short-term missed detections make online identity preservation difficult and often lead to prediction drift, identity switches, and trajectory fragmentation. This paper presents A2MGA-Track, a lightweight online tracker based on motion-geometric association. Without introducing an additional appearance embedding branch into the main association pipeline, A2MGA-Track integrates motion uncertainty-aware prediction (MUP), kinematic-guided discriminative association (KGDA), and reliability-constrained track management (RCTM) into a closed prediction-association-management loop. MUP adapts short-term motion prediction to detection quality and maneuver uncertainty, KGDA refines data association with velocity-geometric context, and RCTM uses track reliability to regulate recovery, retention, and new-track initialization. Experiments on UAVSwarm and MOT-FLY under unified detection inputs show that, among representative tracking-by-detection trackers, A2MGA-Track achieves strong identity preservation with low identity-switch counts while maintaining real-time pipeline efficiency. These results indicate that lightweight motion-geometric association provides a practical alternative for real-time air-to-air multi-UAV tracking under weak appearance and complex relative motion.