LiDAR-based Global Registration Algorithms for Pose Acquisition of Non-cooperative Spacecraft

Abstract

This thesis work, developed in collaboration with the German Aerospace Center (DLR), is placed in the context of spacecraft pose determination, i.e., the problem of calculating the set of parameters that describe the relative position and attitude of an active satellite with respect to another space object, which is widely encountered in space missions such as On-Orbit Servicing (OOS) and Active Debris Removal (ADR), where algorithmic and technological solutions are essential to ensure the efficient execution of autonomous maneuvers of a chaser in close-proximity with respect to a designated target.

Specifically, the work carried out addresses the problem of pose acquisition of a known non-cooperative spacecraft, based on the use of target 3D point cloud scans produced by a LiDAR sensor, proposing a suite of feature-based algorithmic solutions, developed in Python environment, that leverage point-normal structures as local features (Fast Point Feature Histograms, FPFH) or as non-local primitives (Point Pair Features, PPF). Additionally, they exploit a Random-Sample-Consensus-based (RANSAC-based) strategy to perform the initial pose estimation and Hash Tables (HT) for fast and efficient matching.

The performance of the proposed architecture is tested using a dataset of synthetic point clouds obtained using a LiDAR data simulator developed by DLR and considering as target the Client Satellite of the DLR On-Orbit Servicing Simulator for Capture (OOS-SIM). The achieved performance is compared against standard approaches in 3D registration, namely FPFH-based RANSAC and Fast Global Registration (FGR), implemented in the Python Open3D Library. The obtained results demonstrate that these algorithms are promising alternatives to standard approaches, showing comparable accuracy, but with a slight disadvantage in computational time.

Finally, a description of an autonomous failure detection strategy is provided, which can be applied to increase robustness of the proposed pose estimation architectures.