Contextual boundary-aware network for semantic segmentation of complex land transportation point cloud scenes

Abstract

Semantic segmentation of land transportation scenes is critical for infrastructure maintenance and the advancement of intelligent transportation systems. Unlike traditional large-scale scenes, land transportation environments present intricate structural dependencies among infrastructure elements and pronounced class imbalance. To address these challenges, we propose a Gaussian-enhanced positional encoding block that leverages the Gaussian function’s intrinsic smoothing and reweighting properties to project relative positional information into a higher-dimensional space. By fusing this enhanced representation with the original positional encoding, the model gains a more nuanced understanding of spatial dependencies among infrastructures, thereby improving its capacity for semantic segmentation in complex land transportation scenes. Furthermore, we introduce the Multi-Context Interaction Module (MCIM) into the backbone network, varying the number of MCIMs across different network levels to strengthen inter-layer context interactions and mitigate error accumulation. To mitigate class imbalance and excessive object adhesion within the scene, we incorporate a boundary-aware class-balanced (BCB) hybrid loss function. Comprehensive experiments on three distinct land transportation datasets validate the effectiveness of our approach, with comparative analyses against state-of-the-art methods demonstrating its consistent superiority. Specifically, our method attains the highest mIoU (91.8%) and OA (96.7%) on the high-speed rail dataset ExpressRail, the highest mIoU (73.3%) on the traditional railway dataset SNCF, and the highest mF1-score (87.4%) on the urban road dataset Pairs3D. Codes are uploaded at: https://github.com/Kange7/CoBa.