Study On Building Extraction From LiDAR Data

The 3D city model is the foundation of digital city. The building model is the most important part of the 3D city model. The 3D information of buildings in urban area can be acquired accurately, quickly and automatically by LIDAR. The model of the builidngs can then be extracted from the LIDAR data. In this thesis, we studied the extraction of building model from the dense and unorganized 3D LIDAR points. The challenge of modeling LIDAR point cloud is to create organized, simple, noise-free and watertight models of buildings, which can be rendered in real-time. Most exist methods follows a three steps pipeline:first, irrelevant parts such as trees and noise are removed from the LIDAR data; second, background and objects are extracted from the remaining points respectively; finally, the individual object models and background model are generated. In this thesis, we follow the pipeline to extract the 3D model of buildings.The first step of our approach is LIDAR data preprocessing, in which the irrelevant part are removed. The method based on improved covariance analysis is applied in this step to extract the feature of the LIDAR points and remove irrelevant points. The covariance analysis is a widely used method in existed works. In this thesis, it is improved by using different types of neighborhoods. For a given point, five different neighborhoods are used. The covariance analysis is taken on all of them. The fittest neighborhood is chosen. The covariance analysis result of this neighborhood is used to exact the features of the given point. The improved covariance analysis can handle the difficult situation on which the traditional one fails.The second step is building detection, in which the points from buildings and ground are separated. In this step, the classification of ground and buildings is based on the connect components. The traditional classification is achieved by assigning the ground label to the largest connected component, and labeling the rest component as buildings. This method is quick and simple. However, it doesn't work sometimes. When the ground is separated into small pieces due to the occlusion of high buildings and noise, the largest component will not be the ground and the smaller ones will not always be the buildings.In this thesis, a Markov Random Field is modeled on the connected components. For each component, a marginal distribution, which is used to classify this component, is calculated. The LIDAR points, which belong to certain class of component, are labeled as corresponding class.The final step is building and ground modeling. The geometry of the buildings is reconstructed by 2.5D dual contouring. The Digital Surface model of the ground is generated by Laplace interpolation. The texture of both building models and ground model is generated from the visible image, whose acquisition is simultaneous with the acquisition of LIDAR points. In some case, a texture of infrared or other spectral may need. In this thesis, we provide a method to generate such texture by multi-spectral image registration technique.The experiment result shows that the improved covariance analysis can improve the classification of irrelevant points and other points. The classification based on Markov Random Field can correctly classify the points from buildings and ground. The classification is more robust than the traditional method. Also, the 3D model with multi-spectral texture of buildings and ground can be properly reconstructed by the modeling method proposed in this thesis.