3D Model Points Reconstruction and Pose Solver for Vision Based Relative Navigation between Spacecrafts

Abstract

Relative navigation between spacecrafts has been in research for quite a long time but recently it gained importance due to applications in proximity operations and debris removal. Moreover, the use of Artificial Intelligence and Deep Learning in terrestrial applications especially in Robotics has opened new areas of advancement in space applications as well. In this work, two frameworks are presented that are essential part of pose estimation pipeline for vision based relative navigation between chaser and client spacecraft. In first part, 3D model is reconstructed using 2D images of the target satellite from monocular camera. Inverse Direct Linear Transform (IDLT) and Singular Value Decomposition (SVD) methods are used for 3D reconstruction. Secondly, Perspective-n-Point (PnP) problem is solved to estimate the pose of target satellite with respect to the camera on client satellite. PnP estimate the initial guess of pose through closed form analytical approach using Direct Linear Transform (DLT) and then non-linear Gauss Newton Method (GNM) is used over it for best fit of the pose. The results are evaluated on SPEED dataset. Pose errors with number of iterations for both rotation and translation motion are calculated and compared with ground truth values.