Pixel Voting Network-Based Monocular Pose Estimation for Space Non-Cooperative Targets

This paper proposes a monocular pose estimation method for space non-cooperative target based on pixel-wise voting network (PVNet). By combining deep learning with geometric optimization, our approach can effectively improve the accuracy of pose estimation and solve the problems of low computational efficiency caused by train two networks: the object detection network and the key point regression network. In this paper, the 3D key points of the non-cooperative target are reconstructed by a multi-view triangulation, the 2D key point regression is performed using PVNet, and the non-cooperative target's pose is estimated using the Efficient Perspective-n-Point (EPnP) method. Experimental evaluation on a public dataset shows that the method in this paper can estimate the pose of non-cooperative targets efficiently and accurately.