Deep Learning-Based Automation of Road Surface Extraction from UAV-Derived Dense Point Clouds in Large-Scale Environment

Investigating road conditions is an essential task to maintain service life, ensure safety, and preserve smoothness. Digital road surface information is gaining interest as the main database for planning, design, and inspection. However, the implementation of this still confronts several challenges, such as automated surface extraction, typical geometric features, large scene size, and so on. Although active research on point cloud-based automated road surface extraction has addressed this issue, previous studies mainly relied on experts’ perceptions to decide on the result produced by the selected semantic classification and estimated value, consequently requiring trial and error. This study proposes a U-net image segmentation model-aided method to automatically extract 3D point cloud road surface in typical circumstances generated by UAV images. The proposed method starts by developing a road curb detecting model for geometric 3D spatial data with U-net. Then, the road surface is derived by using geometric feature extraction and machine learning classification algorithms. The outcome is assessed by comparing it with manually classified points and the surface produced by a method from a previous study. The proposed method is capable of extracting road surfaces while reducing human involvement, elevating the road extraction from an unstructured 3D point cloud to a more fully automated approach.