Benchmarking motion-based tracking algorithms for small UAV object tracking

Abstract

The growing use of small unmanned aerial vehicles (UAVs) creates serious security challenges for airports, critical infrastructure, and military sites, driving demand for automated counter-drone systems that can detect and track aerial targets in real time. Tracking small UAVs remains hard because of low-resolution appearance, irregular flight dynamics, frequent occlusions, and the high sensitivity of the IoU metric to pixel-level localization errors, which together make reliable temporal association difficult. Most multi-object tracking (MOT) benchmarks focus on pedestrians rather than small aerial targets, leaving practitioners without clear guidance on tracker choice for the UAV domain. This paper addresses that gap with an empirical evaluation of five SORT-family multi-object tracking algorithms, namely ByteTrack, BoT-SORT, DeepOC-SORT, Hybrid-SORT, and StrongSORT, on the Anti-UAV-RGBT dataset. All trackers share an RT-DETR-L detector fine-tuned on the Anti-UAV-RGBT training split, and appearance re-identification is disabled so that the comparison reflects only motion-based association strategies. The test set contains 92 sequences with 79,699 annotations, grouped into three size categories (small ≤48 px, medium 48–128 px, large >128 px) to enable a size-stratified break-down of tracking quality. Experiments run on a single NVIDIA RTX 5090 GPU. BoT-SORT achieves the highest overall MOTA of 0.7802 and also leads on small objects with a MOTA of 0.7479; its Global Motion Compensation module cancels camera-induced displacement before Kalman prediction and preserves IoU overlap when targets span only a few dozen pixels. ByteTrack shows the largest drop on small objects, losing 10.8 percentage points relative to its overall score. DeepOC-SORT and Hybrid-SORT run 38% faster (over 33 FPS) while giving up less than 1.3 percentage points of overall tracking accuracy. Based on these findings, the paper provides practical deployment guidelines that help practitioners pick a tracker according to target scale, accuracy expectations, and available compute budget.

Keywords: UAV tracking, multi-object tracking, SORT, small object tracking, RT-DETR, benchmark, Anti-UAV.