Deep Learning-Based Pose Estimation of Non-Cooperative Spacecraft from Sparse 3D Point Clouds

This paper investigates the use of Deep Learning techniques for LiDAR-based pose initialization of a known, non-cooperative spacecraft. Three point-based neural networks – PointNet (PN), a two-layer PointNet-based encoder (PN-PCN), and PointNet++ (PN++) – and three attitude parametrizations – quaternion, 6D, and soft-label representations – are examined to identify the most suitable architecture for processing sparse point clouds. Moreover, a performance analysis of the selected architecture is conducted including an ICP-based pose refinement step, using synthetic datasets generated with an in-house-developed LiDAR simulator. Results show that, despite the sparsity of the point clouds, the PN-PCN + soft-label configuration achieves a mean angle error of 0.92° and a mean translation error of 8 cm on a dataset with random 30-80 m distances considering the ENVISAT satellite as a target.