Scan-to-graph: Automatic generation and representation of highway geometric digital twins from point cloud data

Abstract

Constructing geometric digital twins of highways at present still demands substantial human effort. Unlike most previous work that uses deep learning models to segment point clouds of highways into class level or object instance level, this paper further segments pavements into a more detailed level (lanes, hard shoulders, central reserves). The central curves of each lane marking are fitted in a two-step method, approximated by a polynomial and then converted into the Frenet coordinated system. The fitted curves with smoothly changing curvature are used to separate points of road surfaces into lanes, hard shoulders, and central reserves, resulting in a mean Intersection over Union (mIoU) at around $90%$. This automatic approach extracts geometric and object category information from point clouds and stores the information in a graph, showing the hierarchical relationships among various components and offering the potential for expansion into more comprehensive digital twins encompassing the entire highway network.