AURORA-Track: Uncertainty-Aware Identity Prediction for Robust Multi-Object Tracking in Satellite Video

Satellite video multi-object tracking remains difficult due to tiny targets, frequent cloud/shadow occlusion, and strong cross-scene domain shifts. To address these issues, we propose AURORA-Track, an end-to-end framework built on MOTIP with three coupled components: uncertainty-aware identity prediction (UAI), cloud/shadow-aware trajectory modeling (CAT), and cross-scene adaptation (CSA). UAI calibrates association confidence to reduce unreliable identity assignments; CAT combines visibility estimation with motion-state recovery to maintain trajectories through prolonged occlusions; CSA uses lightweight FiLM/LoRA adaptation for improved generalization under geographic variation. Experiments on VISO, SAT-MTB, and AIR-MOT-100 show consistent gains in HOTA and IDF1 with fewer identity switches compared with strong baselines. The results indicate that coupling uncertainty estimation, visibility-aware motion reasoning, and scene adaptation improves robustness for satellite video tracking.