Abstract
Detecting and tracking non-cooperative targets in sequential star sensor images poses several challenges in rapid expanding space involving mega constellations, commercial space activities: poor real-time performance in target acquisition, inadequate generalization to different target speeds, and dependence on the star sensor's attitude priors. To address these challenges, an improved multi-target tracking deep learning model based on the CenterTrack model is proposed in this paper. A sophisticated training and testing dataset that closely replicates real on-orbit star sensor images is constructed. By aggregating features across the entire sequence and enhancing the target identification ability from background noise, an improved tracking accuracy is realized. These improvements reduce false positive rates by approximately 60% and lower true target miss rates by 20% compared to the baseline original CenterTrack model. Furthermore, adjusting hyperparameters and optimizing the tracking algorithm reduces target ID switching frequency by approximately 50%. Compared to the traditional algorithm, the improved model can accurately capture newly appearing targets using only two frames, achieving a six-fold speed improvement. The generalization performance is significantly improved with respect to variations in target morphology and velocity, thus a higher target speed tolerance is achieved. The proposed model eliminates the requirement of external attitude priors, thereby enhancing its robustness, and shows significant potential for emerging on-orbit non-cooperative target tracking applications.