Pose Estimation for Non-cooperative Spacecraft based on Deep Learning
Wenxiu Huan, Mingmin Liu, Qinglei Hu
Abstract
The pose estimation of non-cooperative space-borne object is of vital importance for on-orbit service and spacecraft approaching missions. Based on images taken by a monocular camera, an estimate algorithm is proposed to estimate the relative position and the relative attitude of non-cooperative spacecraft. The approach utilizes the off-the-shelf target detection network and key point regression network to predict 2D key points coordinates and combines the multiple view triangulation to reconstruct 3D model. Nonlinear least squares method is used to minimize 2D-3D corresponding coordinates to predict position and attitude. The proposed method effectively combines deep learning and geometric optimization algorithms, which is an innovative application of deep learning in the aerospace field. Finally, simulations are performed to prove the effectiveness of the theoretical results.