Reliability-Guided Tracklet Refinement for UAV Multi-Object Tracking

Abstract

Unmanned Aerial Vehicle (UAV) video streams present challenging conditions for multi-object tracking (MOT) due to dense target distributions, small object sizes at high altitudes, partial occlusions, and simultaneous camera and object motion. Even state-of-the-art online trackers generate temporary association failures; a single physical trajectory may be fragmented into several identities, whereas false detections lead to short false tracklets. To address these challenges, we propose Reliability-Guided Tracklet Refinement (RGTR), a lightweight, training-free post-processing method that directly refines tracker output. RGTR operates in two sequential phases: the Tracklet Reconnection Phase (TRP) and the Tracklet Filter-ing Phase (TFP). TRP bridges reliable trajectory fragments by evaluating a composite reliability cost that integrates temporal and spatial information. In contrast, TFP reduces false positives by eliminating transient tracklets that lack sufficient confidence and temporal continuity. Experiments on the VisDrone2019-MOT benchmark using YOLO11s with BoT-SORT, ByteTrack, and OC-SORT demonstrate that RGTR consistently improves association quality and overall tracking accuracy. On the VisDrone2019-MOT validation split, RGTR improves the BoT-SORT baseline by 4.67, 1.71, and 0.87 percentage points in MOTA, IDF1, and HOTA, respectively, while reducing IDSW by 274 switches (42.4%). On the test-dev split, RGTR achieves average gains of 1.57, 1.48, and 0.61 percentage points in MOTA, IDF1, and HOTA across all three trackers, and reduces IDSW by an average of 695 switches per tracker (24.7%), with an end-to-end runtime overhead of 4.54% to 10.29%. These results confirm that RGTR enhances the reliability of existing online trackers in UAV-based MOT scenarios with minimal computational cost.