Effective 3D Building Extraction from Aerial Point Cloud Data

Abstract

Building extraction is important for a wider range of applications including smart city planning, disaster management, security, and cadastral mapping. This thesis mainly aims to present an effective data-driven strategy for building extraction using aerial Light Detection And Ranging (LiDAR) point cloud data. The LiDAR data provides highly accurate three-dimensional (3D) positional information. Therefore, studies on building extraction using LiDAR data have broadened in scope over time. Outliers, inharmonious input data behaviour, innumerable building structure possibilities, and heterogeneous environments are major challenges that need to be addressed for an effective 3D building extraction using LiDAR data. Outliers can cause the extraction of erroneous roof planes, incorrect boundaries, and over-segmentation of the extracted buildings. Due to the uneven point densities and heterogeneous building structures, small roof parts often remain undetected. Moreover, finding and using a realistic performance metric to evaluate the extracted buildings is another challenge. Inaccurate identification of sharp features, coplanar points, and boundary feature points often causes inaccurate roof plane segmentation and overall 3D outline generation for a building. To address these challenges, first, this thesis proposes a robust variable point neighbourhood estimation method. Considering the specific scanline properties associated with aerial LiDAR data, the proposed method automatically estimates an optimal and realistic neighbourhood for each point to solve the shortcomings of existing fixed neighbourhood methods in uneven or abrupt point densities. Using the estimated variable neighbourhood, a robust z-score and a distance-based outlier factor are calculated for each point in the input data. Based on these two measurements, an effective outlier detection method is proposed which can preserve more than 98% of inliers and remove outliers with better precision than the existing state-of-the-art methods. Then, individual roof planes are extracted in a robust way from the separated outlier free coplanar points based on the M-estimator SAmple Consensus (MSAC) plane-ftting algorithm. The proposed technique is capable of extracting small real roof planes, while avoiding spurious roof planes caused by the remaining outliers, if any. Individual buildings are then extracted precisely by grouping adjacent roof planes into clusters. Next, to assess the extracted buildings and individual roof plane boundaries, a realistic evaluation metric is proposed based on a new robust corner correspondence algorithm. The metric is defined as the average minimum distance davg from the extracted boundary points to their actual corresponding reference lines. It strictly follows the definition of a standard mathematical metric, and addresses the shortcomings of the existing metrics. In addition, during the evaluation, the proposed metric separately identifies the underlap and extralap areas in an extracted building. Furthermore, finding precise 3D feature points (e.g., fold and boundary) is necessary for tracing feature lines to describe a building outline. It is also important for accurate roof plane extraction and for establishing relationships between the correctly extracted planes so as to facilitate a more robust 3D building extraction. Thus, this thesis presents a robust fold feature point extraction method based on the calculated normal of the individual point. Later, a method to extract the feature points representing the boundaries is also developed based on the distance from a point to the calculated mean of its estimated neighbours. In the context of the accuracy evaluation, the proposed methods show more than 90% F1-scores on the generated ground truth data. Finally, machine learning techniques are applied to circumvent the problems (e.g., selecting manual thresholds for different parameters) of existing rule-based approaches for roof feature point extraction and classification. Seven effective geometric and statistical features are calculated for each point to train and test the machine learning classifiers using the appropriate ground truth data. Four primary classes of building roof point cloud are considered, and promising results for each of the classes have been achieved, confirming the competitive performance of the classification over the state-of-the-art techniques. At the end of this thesis, using the classified roof feature points, a more robust plane segmentation algorithm is demonstrated for extracting the roof planes of individual buildings.