Space Target Pose Estimation Framework with Deep Reinforcement Learning Technique

Vision-based relative navigation technology is the key technology to complete on orbit service, space debris removal, and formation flight. One of the current challenges to this technology is to estimate the attitude of uncooperative targets that do not provide any navigation assistance. Previously, the development of vision-based relative navigation technology has mostly relied on image processing and template matching techniques, such as feature-based matching and tracking methods, based on PnP or PnL methods. Line features were often adopted for pose estimation in important on-orbit experiments since the advantages of robust detection and more structural information and less likely to be affected by occlusions. When the corresponding relationship between the model lines and the image projection lines is known, an accurate relative pose could be solved by Newton and Levenberg-Marquardt methods. For unknown line correspondences case, this problem is difficult to solve. This paper explores a pose estimation framework that uses deep reinforcement learning to learn the matching relationship and relative pose. The Deep Deterministic Policy Gradient (DDPG) algorithm is adopted since it supports the continuous action space and continuous status space. While the MDP is delicated designed, with the rotations angle being included in the action space, through continuous attempts and exploration, the model would finally reaches the "correct" posture, and create a match between the model lines and image lines.