Reliability-Aware Spatiotemporal Consistency Regulation for Satellite Video Tracking

Abstract

The satellite video single object tracking (SVSOT) remains challenging because low image quality and cluttered backgrounds yield unreliable target evidence. Although recent trackers have improved robustness by enhancing feature representations and cross-frame interactions, they lack regulation of the reliability of evidence used for recursive feature updates. As single object tracking (SOT) can be seen as a recursive process, once unreliable evidence is absorbed into the target state, it may bias subsequent associations, accumulate spatiotemporal inconsistency, and eventually cause tracking drift or failure. To tackle these issues, we propose SCRTrack, a spatiotemporal consistency regulation framework that regulates the formation of target evidence and its propagation and maintenance across frames. Specifically, a multilayer feature formation (MFF) module is proposed to inject shallow structural details into deep semantic representations, enhancing spatial discriminability while suppressing clutter-induced interference. A reliability-guided context propagation (RCP) module is further designed to estimate interframe correspondence reliability and uses it to modulate temporal state updates, reducing the impact of ambiguous associations on state evolution. In addition, a motion consistency preservation (MCP) mechanism is introduced to monitor motion-consistent response displacements under occlusion and similar-object interference, maintaining target identity when appearance-based evidence becomes unreliable. Extensive experiments on SatSOT, SV248S, and OOTB demonstrate that SCRTrack achieves the state-of-the-art performance and exhibits strong robustness across challenging satellite tracking scenarios.