SegRail: A Framework for semantic segmentation of long-distance railway MLS data via global priors and local geometry integration

Mobile Laser Scanning (MLS) enables accurate and efficient data acquisition in long-distance railway environments, yet directly processing the resulting large-scale point clouds imposes a prohibitive computational burden. Conventional rigid discretization methods induce artificial boundary artifacts and disrupt the intrinsic topology of continuous assets, compromising the structural integrity required for reliable engineering monitoring. To address these challenges, we propose SegRail, a dedicated framework for semantic segmentation of long-distance railway MLS data that explicitly integrates global priors and local geometry. The framework first introduces a sectioning strategy that converts continuous railway scenes into structured segments, reducing computational complexity while preserving geometric coherence. Based on these segments, we design a semantic segmentation network that jointly models global semantic priors and local geometric structures. Specifically, a global prior extraction module is introduced to capture long-range contextual dependencies and category distribution priors, while a specialized neighborhood operator encodes fine-grained local geometric relationships to preserve structural boundaries. Extensive experiments on five railway datasets demonstrate the effectiveness of SegRail, which achieves the highest mIoU than the compared baseline methods under multiple runs. Specifically, it achieves mIoUs of 84.37%, 77.36%, 75.35%, 83.55%, and 92.60% on WhuRailway3D-Urban, WhuRailway3D-Rural, WhuRailway3D-Plateau, SNCF, and ExpressRail, respectively.