An End-to-End Pose Estimation Network for Multiscale Space Non-cooperative Objects

The pose estimation method of space non-cooperative objects based on deep learning is of great significance for further improving the on-orbit service level. However, at present, the pose estimation method of space non-cooperative objects based on deep learning still has some problems, such as relying on a priori 3D wireframe model, the network is not lightweight enough, and the precision of pose estimation for multiscale objects is not high. Therefore, this paper proposes an end-to-end pose estimation network for multiscale space non-cooperative objects. First, the lightweight EfficientNet-B0 is selected as the backbone, and feature pyramid network is introduced into EfficientNet-B0 to improve the pose estimation precision of the network for non-cooperative objects at middle and far distance. Then, a pose prediction head network including object loss function and pose loss function is designed. Finally, a lightweight and multiscale pose estimation network for space non-cooperative objects is established. The simulation results in SwissCube dataset show that the proposed pose estimation network has an average precision improvement of 3.6% compared with advanced methods. In addition, compared with other backbones, the “EfficientNet-B0+FPN” improves the average precision by 7.6% and is lighter.