Multi‐Object Tracking of 3D faint space objects via cascade detection and transformer based motion modeling
Haotian Wang, Tao Zhan, Xiaotong Zhu, Zirui Li, Shuqi Hou, Shixiang Cao, Mengjie Hu · Space habitation. · 2025
Tracking faint objects in deep space poses three major challenges in aerospace exploration: low signal-to-noise ratio (SNR), motion noise interference, and multi-view matching. This paper proposes an integrated solution combining detection, tracking, and 3D reconstruction. First, a cascade detection algorithm is designed to achieve efficient detection of low-SNR objects (1.5–6) through global feature enhancement and local background suppression. Second, a hybrid tracking framework is developed, integrating Kalman filter for observation error compensation and Transformer networks for long-term motion modeling, effectively addressing tracking interruptions caused by temporary object disappearance. To our knowledge, this is the first integration of Kalman smoothing and transformer-based motion modeling applied specifically to low-SNR deep space tracking scenarios. Finally, a multi-view object matching strategy is established, optimizing the Hungarian algorithm with epipolar constraints to associate objects in binocular systems, followed by 3D coordinate calculation for spatial trajectory reconstruction. Simulation experiments demonstrate that the proposed method significantly outperforms traditional approaches in object localization accuracy, trajectory continuity, and 3D reconstruction error, providing an effective technical pathway for faint object tracking in deep space exploration.