MOSAIC-Tracker: Mutual-enhanced Occlusion-aware Spatiotemporal Adaptive Identity Consistency network for aerial multi-object tracking

Abstract

Multi-Object Tracking (MOT) in aerial imagery remains challenging due to small object sizes, occlusions, and dynamic environments. Existing approaches predominantly rely on high precision detection and Re ID matching but neglect spatiotemporal cues and global temporal modeling of occlusion. Their static confidence weighting during association cannot adapt to real time detector confidence fluctuations, resulting in mismatches and ID switches. To alleviate these limitations, we propose MOSAIC-Tracker, a Mutual-enhanced Occlusion-aware Spatiotemporal Adaptive Identity Conservation Network with three key dimensions. First, a Spatiotemporal Occlusion Enhancement (STOE) module integrates multi-frame temporal dependencies to model global motion patterns and local dynamic features, mitigating identity switches during occlusions. Then, an Adaptive Multi-scale Feature Enhancement (AMFE) mechanism combines a Local Enhancement Mechanism with multi-scale feature aggregation to improve small object discrimination. Finally, a Dynamic Confidence Matrix Adjustment (DCMA) strategy adaptively weights detection confidence in trajectory matching to minimize association errors. Together, the three modules reduce occlusion-induced identity switches. Extensive evaluations on UAVDT and VisDrone2019 datasets demonstrate advanced performance. The code is released at: https://github.com/aJanm/MOSAIC-Tracker.