HiMoTrack: Small-Object Perception and Hierarchical Motion Recovery for UAV Multi-Object Tracking
Abstract
Unmanned aerial vehicle (UAV)-based multi-object tracking (MOT) is important for intelligent traffic analysis but remains challenging because small object scales, frequent occlusions, motion blur, and changing aerial viewpoints jointly degrade detection quality and trajectory continuity. This paper presents HiMoTrack, a tracking-by-detection framework that addresses this degradation chain by improving small-object observations and progressively recovering low-quality tracks. Specifically, a Small Object Feature Efficient Perception Network (SOFEPNet) preserves spatial detail and reconstructs multi-scale features for aerial small objects. For motion recovery, a Velocity Adaptive Cubature Kalman Filter (VACKF) provides reliability-adaptive short-term state estimation, while a Group Motion Compensation Strategy (GMCS) exploits local group-motion consistency when individual observations become unreliable. When no reliable neighboring motion prior is available, Spatio-Temporal Memory Prediction (STMP) uses historical trajectory memory to infer the missing state and maintain temporal continuity. In this way, the detector, motion estimator, and trajectory recovery process are organized around the same objective: stabilizing weak and interrupted trajectories in complex UAV traffic scenes. Experiments show that HiMoTrack achieves multi-object tracking accuracy (MOTA) scores of 64.6% on the UAVDT benchmark and 62.4% on VisDrone while maintaining strong overall tracking accuracy and competitive trajectory continuity. Additional experiments on MOT17 demonstrate that the proposed motion recovery strategy can be integrated into representative trackers, indicating good modularity and transferability.