Abstract: To address the challenges of extracting the stereoscopic road network structure from large-scale point cloud scenes, this study proposes a method for 3D point cloud road network extraction based on L₁ medial axis extraction and flexible constraints. The method begins by using the L₁ skeleton extraction algorithm to obtain the initial 3D skeleton. Next, a centroid distance field is constructed using a symmetry method, and the ground point cloud is eroded to constrain the road’s medial axis area. Subsequently, the endpoints of the initial road network skeleton are evaluated, and the network is completed accordingly. Finally, an optimal medial axis position for the 3D road point cloud is achieved through flexible projection. Experiments on large-scale scene datasets Campus3D, UrbanScene3D, and UrbanBIS show that the proposed method can achieve a completeness of up to 95.99% within a 0.4 m range, with errors less than 0.2 m in the first 88% of the data. Compared to the 2D road network extraction method MTH, it effectively captures the road network’s 3D structure.