Abstract: Road boundaries and markings, as crucial linear elements of road surfaces, provide vital support for highway expansion, reconstruction, and intelligent asset management by offering accurate positioning and semantic information. Mobile LiDAR scanning (MLS) point clouds have emerged as a significant data source for identifying and updating road scene elements. Addressing shortcomings of existing methods in complex mountainous highway scenarios, this study proposes a progressive method for intelligent extraction of linear road surface features from vehicle-mounted LiDAR point clouds: Utilizing a deep learning network based on 3D point clouds, semantic segmentation of road surfaces and markings is performed to obtain coarse results. Subsequently, refined extraction and optimization processes are applied to road boundaries and markings. This includes extraction of road contour points, clustering of line structures, completion and vectorization of road boundary points, and semantic enhancement through feature attribute analysis and edge-based feature matching for precise extraction and semantic interpretation of markings. Experimental results demonstrate that the proposed method effectively addresses occlusion and noise challenges in complex mountainous highway scenes, achieving accurate extraction of linear road surface features.

Key words: MLS point cloud, semantic segmentation network, highway pavement, lane marking, edge matching