PF-Net++: high-fidelity completion of roadway point cloud from LiDAR sensor via combined structure-aware point transformer and multiscale skip connections

This study aims to address the challenge of incomplete point clouds acquired by underground LiDAR sensors under dust, water mist and occlusion conditions. This study focuses on achieving high-fidelity 3D measurement and completion of coal-mine roadways to ensure sensing accuracy and measurement reliability for intelligent mining. The authors propose PF-Net++, an enhanced point fractal network integrating a combined structure-aware point transformer (CSA-PT) encoder with a point-pyramid decoder using multiscale skip connections. The CSA-PT leverages the roadway centerline as a structural prior to guide anisotropic neighborhood construction. In addition, the authors construct a real-world benchmark, termed the roadway point cloud patch data set (RPCD), for model evaluation. Experiments on the RPCD show that PF-Net++ achieves the best overall performance among the compared baselines across most evaluation metrics. Compared to SymmCompletion, it achieves a 32.3% reduction in Chamfer distance (CD) (GT→Pred), a 6.3% gain in the strictest F1-score (d = 0.01), and the lowest CD (Pred→GT) of 0.3508. Evaluations on Point Completion Network and multi-view partial benchmarks further demonstrate competitive generalization beyond roadway scenes. Ablation studies further confirm the complementarity of the proposed components, yielding a 25.2% improvement in F1-score and a 19.8% reduction in CD compared with single-component variants. This research integrates a roadway centerline structural prior into a CSA-PT to handle weakly textured environments and contributes the RPCD benchmark. The proposed framework enables high-fidelity completion of roadway point clouds, thereby improving data quality for underground digital mapping and safety assessment.