Multi-Target UAV Tracking with Feature Optimization and ID-Level Contrastive Learning

To tackle challenges in UAV-based multi-object tracking (MOT) such as large target scale variations, frequent occlusions, and unstable appearance features, this paper proposes UAV-AttendNet, an end-to-end unsupervised tracking framework. It comprises three synergistic modules: an attention-enhanced feature encoder, a trajectory correction module, and a contrastive learning module with hard negative sample weighting. The encoder applies spatial and channel attention to multi-scale features, improving discrimination for occluded targets. The trajectory correction module leverages pseudo-label confidence to refine associations and enhance tracking quality. The contrastive learning module employs a similarity-driven hard negative weighting strategy to strengthen embedding discriminability. Experiments on UAVDT and VisDrone-MOT datasets demonstrate that UAV-AttendNet outperforms existing methods, achieving superior results in unsupervised MOT tasks.