Pose Estimation of Non-Cooperative Spacecraft from LiDAR: Comparison of Methods Based on Point Correspondences

This paper deals with LiDAR-based pose estimation of a known, non-cooperative spacecraft in mission scenarios requiring close-proximity robotics operations, like On-Orbit Servicing and Active Debris Removal. Attention is focused on classical methods, which today are still relevant in the space domain, being the only solution for safety-critical real-time applications that is space-qualifiable with recognized standards. In this work, different classical algorithms for global pose estimation are compared. We study variants of methods, both known and new ones, which support RANSAC search by guiding point correspondences between measured and model point clouds, using point-normal structures as local features (FPFH) or as non-local primitives (PPF). We also compare these methods to Fast Global Registration (FGR) which treats global registration as a robust optimization problem without assigning hard correspondences. For all the approaches, after the global pose initialization, the estimate is refined using the point-to-plane Iterative Closest Point (ICP) algorithm. The pose estimation pipeline also includes additional evaluation steps to resolve pose ambiguities. All methods are tested using a dataset of synthetic partial and sparse point clouds obtained with a LiDAR data simulator. The methods based on point correspondences outperform FGR by a large margin. Excluding FGR, they are all promising candidates for the considered application with some slight differences in speed and accuracy.