Cross-Scale Point Cloud Completion Method for Space Non-cooperative Targets Based on Probability Distribution Generation

In space debris capture missions, the missing or incomplete point cloud data caused by sensor characteristics and occlusion factors increase the difficulty of subsequent recognition tasks. Point cloud completion can enhance its three-dimensional structural information to improve the accuracy of recognition tasks. However, the structural scale differences of space non- cooperative targets are significant, and the edge loss of point clouds collected in a rolling state is severe. Existing point cloud completion methods overlook the impact of scale differences on point clouds, making it challenging to obtain diverse information on local geometric details under cross-scale conditions. To address these issues, we propose a novel two-stage dual-channel completion network, PSPC, to generate complete point clouds. By combining skeleton point features with probability distribution features to capture the topological relationships between neighboring regions and distant points in sparse and dense point clouds, we transform these relationships into feature representations that conform to actual distributions to predict missing points with the same probability distribution as the original point cloud. Particularly, we introduce a cross-scale generation module, using the fused features as guidance to adaptively generate rough point clouds of different scales, enabling the completion of targets with different scale point clouds. We conducted extensive experiments on standard and scale- compressed datasets. Qualitative and quantitative experimental results demonstrate that our proposed network outperforms current mainstream methods in cross-scale completion.