Physics-Prior-Driven Feature Enhancement and Semantic Segmentation of High-Speed Railway Line-Scan Point Clouds

Semantic segmentation of railway point clouds is complicated by anisotropic sampling, large data volumes, and small components. This study develops an acquisition-aware processing and segmentation framework that combines structure-aware filtering (SCAF), anisotropy-aware normal estimation, and a modified PointNet++ with reduced-input sampling and component-scale multi-scale grouping. Experiments use a railway dataset containing approximately 400 million raw points. All reported experimental means are based on five independent runs, and run-to-run variation remained within 5% of the corresponding configuration-specific mean. The mixed-normal configuration achieves a mean test mIoU of 97.60% and an overall accuracy of 99.12%. Among the external models, Point Transformer v2 achieves the highest reported mIoU and OA, at 99.10% and 99.80%, respectively. SCAF and statistical outlier removal achieve mean test mIoUs of 95.182% and 95.203%, while their denoising times are 43 s and 18,797 s, respectively. The results indicate that acquisition-aware geometric processing improves component-level segmentation within the investigated setting, whereas the external-model comparison provides a stronger architectural reference for future label-independent evaluation. Validation remains limited to the investigated acquisition system and railway scenario, and the measured processing costs do not establish end-to-end real-time inspection.