Non-cooperative satellite pose estimation based on CPMs

The realization of accurate pose estimation of satellite targets has emerged as one of the research goals as human attention to space exploration has increased. In this paper, we suggest a more accurate estimation technique for non-satellite targets in space. We use the pose machine to extract the key points of the image to form an initialized pose matrix using point pair relationships and initialize the point cloud alignment. First, we extract the pixel coordinates of significant spots from satellite pictures. And then, a rough estimation of the satellite pose at this precise moment is achieved by using the mapping relationship between the picture and real spatial points. To provide an accurate result, the final pose is refined using point cloud data. Our method provides a fast initial attitude input using image features, avoids subsequent point cloud iterative alignment from falling into local optimum, and resolves the satellite attitude ambiguity due to symmetry. Our results show that the method can obtain satellite poses accurately and quickly.