A New Confidence-Aware Adaptive Multi-Object Tracking for Unmanned Aerial Vehicles

Abstract

: Multi-object tracking (MOT) for unmanned aerial vehicles (UAVs) remains challenging due to rapid camera motion, frequent occlusion, scale variation, and unreliable detections, all of which degrade the robustness of data association. Existing detection-based tracking methods generally rely on fixed fusion strategies while lacking explicit reliability modeling for detector observations and motion predictions, limiting their adaptability in dynamic UAV environments. To address this issue, this paper proposes a confidence-aware adaptive data association model for UAV multi-object tracking. The proposed model explicitly characterizes the reliability of detector observations and motion predictions within a unified association framework. A hierarchical confidence model is first developed to progressively associate detections according to their confidence levels. Subsequently, a prediction confidence model is established by exploiting the inverse covariance of the Kalman filter to quantify motion uncertainty. Finally, an adaptive association cost model is formulated by jointly incorporating detection and prediction confidence, enabling dynamic fusion of motion and appearance cues during data association. Experiments on the VisDrone2019 and UAVDT benchmarks validate the effectiveness of the proposed model. The proposed approach achieves 41.439 Multiple Object Tracking Accuracy (MOTA) and 44.001 Higher Order Tracking Accuracy (HOTA) on VisDrone2019, together with 49.000 MOTA and 52.03 HOTA on UAVDT. The results demonstrate that the proposed model provides stable and competitive tracking performance in challenging UAV scenarios.