3D Reconstruction of Buildings from LiDAR Data

This paper presents a new approach to automatic 3D building reconstruction from LiDAR data. While traditional approaches use random sampling or Hugh transform for extracting subsets of coplanar points from noisy point clouds, our method is based on locally fitted surfaces (LoFS). These are planes, best-fitted to the Kneighbourhood of each LiDAR point. In this way, a set of candidate patches for a building surface is obtained. The clustering of patches is then performed based on the planes’ normals and the positions of neighbourhoods, in order to obtain a rough approximation of flat roof sides. An adjacent graph is generated between them and intersections between neighbouring sides are estimated in order to define ridges, while intersections between buildings and ground points are considered in footprint definition. This defines the vertical walls. This method was tested on buildings of different architectural styles, sizes, and complexity. Most buildings are successfully reconstructed, however with increased building details, the accuracy of reconstruction is often decreased.