MDA-Tracker: Motion-Direction Aware Cascaded Association for UAV Multiobject Tracking

Abstract

Multi-object tracking (MOT) in UAV videos is a challenging task because small targets, partial occlusion, bounding-box displacement, and fluctuating detection confidence can reduce the reliability of frame-to-frame data association. In particular, potentially correct track-detection pairs may be rejected when a strict single-stage matching threshold is used. To alleviate this problem, this paper proposes MDA-Tracker, a motion-direction-aware cascaded association method. Its core component, Motion-Direction Aware Cascaded Association (MDACA), is designed as one unified two-stage association mechanism. The first stage applies a confidence-dependent modulation to the association cost, giving reliable detections a moderate favorable bias under a strict threshold. The second stage reevaluates only the unmatched track-detection pairs under a relaxed spatial threshold and introduces motion-direction consistency to suppress candidates that disagree with the predicted short-term motion trend. Experiments on the UAVDT dataset show that MDA-Tracker improves MOTA from 53.6% to 56.0% and IDF1 from 69.8% to 70.8%, while reducing FP by 1742 and FN by 7466. These results indicate that the proposed module mainly improves overall tracking accuracy by reducing false positives and missed associations, although identity switches remain a limitation in dense scenes and under camera motion.