DLA-MOT: A Deep Layer Aggregation Framework for Robust Multi-Object Tracking in UAV Image

To meet the stringent demands of robust and real-time multi-object tracking (MOT) for unmanned-aerial-vehicle (UAV) swarms operating in complex battlefields, we design DLA-MOT, a joint detection-and-tracking framework that synergistically integrates multi-task learning with deep-layer-aggregation. A tree-structured DLA network is designed to iteratively fuse cross-scale semantic and fine-grained features via dense skip connections, yielding a 2.1× gain in representation quality for objects under 32 × 32 pixels. Complementarily, a weighted multi-task loss simultaneously supervises heat-map regression, bounding-box refinement and identity discrimination, enabling balanced co-optimization of detection and re-identification objectives. On top of these, a three-level data-association cascade progressively resolves occlusion, mis-detection and identity recovery by fusing appearance-motion distance, IoU matching and trajectory-confidence gating. Extensive evaluation on the VisDrone-MOT benchmark shows that DLA-MOT attains 30.3 % mAP and 37.3 % mAR, surpassing YOLOX-M, YOLOv11-M and RT-DETR-L by 13.4 %, 13.2 % and 15.5 %, respectively. Tracking performance reaches 47.5 % IDF1 and 32.5 % MOTA while reducing false positives, false negatives and identity switches by 11.3 %, 10.6 % and 32.9 % against the baseline. These results corroborate the framework's robustness and real-time capability across multi-scale, multi-illumination UAV scenarios.