Grayscale Voxel Model Based Airborne LiDAR 3D Road Extraction

2D grid, TIN and point cloud, which are the commonly used methods to represent LiDAR data for road extraction, have defects, for example, it is difficult for 2D grid and TIN to represent multiple return LiDAR data and thus influences the integrity of grid and TIN-based road extraction results and their extraction results are 2D, it is difficult for point cloud to use its topological and adjacent information and thus leads to the difficulty in the design of point-based road extraction algorithm. To overcome these restrictions, a grayscale voxel model (GVM) based 3D road extraction algorithm is presented. LiDAR data are regularized into GVM in which the grayscale of a voxel corresponds to the quantized mean intensity of the LIDAR points within the voxel. Road seed voxels are selected and then seeds and their 3D connected regions are labeled as road voxels. The extracted road result is optimized using mathematical morphology. ISPRS urban LiDAR datasets, which are representative of road networks of different complexities, are used to analyze the sensitivity of "adjacency size" and "intensity difference threshold" parameters and assess the accuracy of the proposed algorithm quantitatively. The experiment results indicate that: 1) 56-adjacency is the optimal adjacency size and 2 is the optimal intensity difference threshold; 2) The average quality, completeness and correctness of road extraction were 70%, 86.77% and 81.13%, respectively; 3) Roads in the relatively flat single layer road network and the undulating complex road network can both be successfully extracted. Copyright © 2020 Acta Automatica Sinica. All rights reserved.