Monocular pose estimation for relative navigation of uncooperative spacecraft

Abstract

This thesis develops a monocular pose estimation framework for the relative navigation of known but uncooperative spacecraft. The objective of relative navigation is to estimate a relative pose that refers to relative position and attitude. This thesis consists of two parts

pose initialization, which estimates the initial pose without pose information in advance, and pose tracking, which generates the relative pose solution based on the initial pose. The first part of the thesis proposes a novel pose initialization algorithm that exploits convexity defects as visual cues. This algorithm selects the neighboring points of the convexity defect in the image and narrows down the matching candidate on the 3D model to find 3D-2D point correspondences. Then, relative poses are calculated from candidate correspondences, and the pose that can produce the most similar image to the given image is determined as the final pose. The final pose is re-examined in the last step of pose initialization to verify if the pose is reliable enough to finish the initialization process. The proposed algorithm is applied to the target spacecraft with a single solar panel on one side for evaluating the estimation accuracy and initialization performance. The second part of the thesis develops Multiplicative Extended Kalman Filter for pose tracking based on monocular images. The loosely-coupled and tightly-coupled approaches are implemented with the synthetic images, showing that the tightly-coupled approach has better pose estimation accuracy. In addition, this thesis proposes a measurement covariance estimation method that reflects the measured relative pose's quality. The modified covariance is applied to the loosely-coupled pose tracking filter, and the simulation demonstrates that the designed covariance estimation method improves the pose estimation accuracy.