Convolutional long-short term memory network for space debris detection and tracking

Space debris detection, as space environments become increasingly crowded, has gradually become a field in urgent need of research. Due to the dim and undersized characteristics of space debris and other limitations in the observation, it often requires combining information from multiple frames to identify desired targets. At present, there are mature model-driven methods available for detecting space targets. In practical applications, however, these methods often encounter issues requiring targeted parameter tuning. Therefore, we adopt a deep learning approach to solve these issues. By employing a recurrent neural network model, we merge target detection and tracking into a unified task, thereby maximizing the utilization of temporal information. Besides, we have implemented backbone network designs and resampling methods tailored for dim and small targets. In the experiments, the proposed method achieved an average recall rate of over 96.3% for targets with SNR over 1.5 in the synthetic dataset with an average frame processing time of 34 ms. Furthermore, a recall rate of 98.1% is obtained when processing the optical images acquired from a wide-field diameter telescope located in Changchun Observatory. These results suggest that the proposed method holds promise for further applications in spatial situational awareness.