LiDAR-based Pose Estimation for Satellites using Point Completion Methods

Abstract

On-orbit servicing is a category of space missions, that aims to increase the life span of orbital satellites and reduce the growing amount of space debris in orbit. For this, servicer satellites are necessary that can interact with orbital target satellites. Such servicers need to first approach their targets for interaction to be possible. This thesis aims to investigate LiDAR-based methods for initial pose estimation of typical target satellites with symmetries, during the approach. The focus is placed on methods using a Point Completion Net architecture for processing LiDAR scans representing a partial geometry of the target. Point clouds that are to be predicted shall have improved expressiveness by representing a completed geometry of the target, by representing a reduced set of feature points or by indirectly implying a pose. Suitable pose derivation methods are used on each of the predicted point clouds. As space missions typically have severe computational limitations, special attention is placed upon the computational complexity of all approaches. The capability of the network architecture to predict point clouds with improved expressiveness and the quality of the derived poses is analyzed systematically for all approaches. The overall potential of one-shot pose estimation methods using Point Completion Net architectures are researched and possible optimizations in consideration of computational complexity and symmetry assumptions have been revealed.