Research on Non-Cooperative Space Satellite Target Position Estimation Based on HRNet Keypoint Extraction

Non-cooperative space target position estimation has become a focus of research since human exploration of space has left a lot of space trash that poses safety issues. Compared to previous methods, deep learning-based approaches are more suited for scenes with multiple targets, complicated backgrounds, and intense lighting. This research proposes a position estimate technique for non-cooperative spacecraft in orbit, based on the HRNet network. The technique extracts feature points from the picture using a heat map, initializes the point cloud using the point-to-point connection between the feature points to generate the initial rotation matrix, and uses the point cloud alignment to estimate its position with accuracy. The pixel coordinates of important spots in a 2D picture are extracted in this article using the HRNet network. To estimate the satellite's approximate position, a mapping connection between the 2D coordinate data and the 3D model coordinates of the relevant spots is used. Point cloud alignment yields precise results for position estimation. In order to address the instability of feature extraction based on the conventional method and prevent position-matching error in the subsequent point cloud alignment, the method in this paper uses neural networks to extract the satellite's key points, obtains a better initial rotation matrix, and applies it for point cloud alignment. In this paper, the satellite's position accuracy in the simulation experiment is measured to be 2 cm, and its angle accuracy is measured to be 1 degree.