PPTracker: Tracking UAV Swarms with Prior Prompt

Haolin Qin Qin, Tianhao Li, Tingfa Xu, Jingxuan Xu, Yuqiang Fang, Jianan Li; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2025, pp. 6594-6601

With commercial drones rapidly gaining popularity, anti-UAV technology is critical to protecting citizen privacy and security. However, there are still many challenges in tracking drones, especially drone swarms, including high inter-target similarity, dense spatial distribution with frequent occlusions, and dynamic scale variations. To overcome these challenges, we introduce PPTracker, a novel Prior Prompt Track framework designed to track UAV swarms in anti-UAV systems. Specifically, PPTracker integrates a detection head based on YOLOv11 and a tracking head utilizing Bot-SORT, enhanced by a dynamic prior prompt encoder. The prompt encoder integrates historical target positions as spatial prior knowledge, employing attention-guided feature refinement to suppress background noise and enhance robustness. The detection head employs the latest YOLO detection framework, providing accurate detection results with high inference efficiency. The tracking head combines motion prediction via Kalman filtering, camera motion compensation, and hybrid appearance-spatial metrics to maintain identity consistency across frames. Evaluated on the 4th Anti-UAV Competition MOT dataset, PPTracker achieves state-of-the-art performance with a MOTA score of 67.9%, significantly outperforming baseline configurations. The framework's effectiveness in handling occlusions and preserving identity coherence is further validated through qualitative visualizations.