Ultra-lightweight ByteTrack-YOLOX-S for UAV-Based Vehicle Tracking on UAVDT

Abstract

Unmanned aerial vehicle (UAV)traffic videos contain small targets, frequent occlusion,moving-camera effects, and large confidence variations, which make lightweight multi-object tracking difficult. This short communication evaluates the authors’ existing YOLOX-S detector and ByteTrack association pipeline on UAVDT under a 10 GB VRAM constraint; it does not introduce a new detector or a new association algorithm. Across 20 evaluated files, 16,583 frames, and 340,906 ground-truth object instances, the training protocol produced 74.8% IDF1, 62.7% recall, 97.4% precision, and 61.0% MOTA, whereas the test protocol produced 48.0% IDF1, 37.2% recall, 78.9% precision, and 27.2% MOTA. On the test protocol, false negatives account for 86.23% of the errors counted by MOTA,while identity switches account for only 0.052%; consequently, the low switch count mainly reflects limited detector coverage rather than consistently strong identity preservation. The revised analysis further explains the two-confidence threshold mechanism, the train-totest degradation, the algorithmic complexity, and the limitations of inferring edge-device latency from desktop-GPU experiments. The results position ByteTrack–YOLOX-S as a compact and transparent baseline whose most important improvement direction is small-object candidate recall under altitude, illumination, and occlusion changes.