THG-Track: Temporal Heterogeneous Graph Structure Learning for Satellite-Video Multi-Object Tracking
Abstract
Multi-object tracking (MOT) in satellite videos is particularly challenging because targets are tiny and crowded, visual contrast is low, and platform-induced motion severely degrades both appearance and motion cues. Existing approaches commonly depend on fixed association heuristics or predefined graph structures, while rarely performing uncertainty-aware fusion of temporal and spatial information. To address these limitations, we propose Temporal Heterogeneous Graph Structure Learning for Multi-Object Tracking (THG-Track), a unified framework that formulates MOT as a temporal heterogeneous graph structure learning problem. Specifically, the proposed method builds a temporal heterogeneous graph over object nodes and coarse prior nodes, and uses graph structure learning to adaptively select and weight typed intra-frame and one-step temporal edges. Based on this graph, relation-specific message passing is performed together with an entropy-gated temporal-spatial fusion mechanism, enabling more reliable association under ambiguous conditions. In addition, the learned graph decisions are fed back to the feature map to benefit detection and association. Extensive experiments on multiple public satellite-video MOT benchmarks show that THG-Track achieves state-of-the-art results and consistently improves tracking performance across diverse scenes and object categories.