Methods Of Regularized Building Models Reconstruction From Aerial LiDAR Data

The development of computer graphics and geospatial technologies leads to the fast progress in the implementation and applications of the Digital City, which has demonstrated its potential as a replica of the real one in urban and environmental planning, design, construction, and management. Buildings are the dominant type of manmade objects in urban scenes, and become the main research focus for photogrammetric and computer vision community. Traditionally, the reconstruction of building models is performed by using high-resolution aerial images, and great development and renovation has been made both theoretically and technically. Although research has progressed significantly, the reconstruction problem is still challenging. Alternatively, the development of Light Detection and Ranging (LiDAR) technology, which is a remote sensing method that uses light in the form of a pulsed laser to measure ranges to the Earth, introduces a better option to reconstruct building models. When compared with the other methods of generating digital elevation data, LiDAR offers several advantages such as high data accuracy, fast data acquisition, high point density, canopy penetration and the minimum of ground truth. Most researches are focused on improving the geometric precision of generated building models and the degree of automation of reconstruction procedures using LiDAR data. The disadvantage of these researches, however, lies in the fact that the generated building models are not regularized enough to practice. Another disadvantage is that existing methods are always self-contained, and it is difficult to introduce external knowledge to improve modelling accuracy and regularity of the complex buildings. Consequently, the aim of this thesis is to propose a data-driven reconstruction method in urban areas, which can generate regularized building models automatically in simple scenes and in a semi-automatic manner in complex scenes.In this thesis, we deeply investigate the key technology for reconstructing regularized building models from LiDAR data. The major works are listed below:1. It’s difficult to preserve terrain features, such as escarpments, terraces or terrain break lines, in filtering LiDAR data in mountain areas. According to the characteristics of LiDAR data in mountain areas, three important aspects of TIN filtering algorithm are improved:more seed ground points are selected using a random strategy; a new ground criteria called predicated terrain angle is proposed for preserving terrain break lines; attaching a priority to each point and adjusting it during filtering, then points with higher priorities will be judged earlier.2. Studies on automatic detection of building information from raw LiDAR data have been raised in recent years. However, it is still difficult to extract building information automatically in complex scenes, because the appearances of trees and buildings in the LiDAR data might be similar. In this thesis, a novel building detection method based on tensor voting theory is proposed. Firstly, the smooth feature is computed using the covariance matrix, then every point in LiDAR data can be encoded as second order, symmetric and positive semi-define3×3tensor. Secondly, a process called adaptive normalization is used to solve the tensor accumulation problem in the voting stage, then an iterative tensor communication scheme, which allows a token to exchange information with its neighbors, and infer new information, is developed.3. The extraction of building boundary is a crucial and difficult step toward generating3D building models. In this thesis, a building boundary extraction algorithm based on semantics is proposed. Firstly, the coarse building boundary is extracted by a dual-threshold alpha shapes method, then simplified by a greedy least-squared method. Secondly, a semantic function is devised to describe the regularity of the building boundary, afterwards the genetic algorithm is used to refine coarse building boundary.4. Two types of approach called region-growing and random sample consensus approaches in the literature are proposed for automatic building roofs extraction. They are, however, constrained by their drawbacks. In this thesis, a hybridized method is proposed to take advantage of both algorithms’ strengths so that building roofs can be extracted more precisely and more efficiently. Firstly, we calculate the normal features and roughness features from LiDAR data for coarse extraction of building roofs. Secondly, precise roof-extraction is performed using an extended RANSAC method which takes the coarse extraction result as the priori-knowledge, and integrates the region growing method. Finally, a poll strategy is adopted to solve the competition problem.5. Automatic reconstruction of buildings from LiDAR data has been regarded as a very difficult problem. In this thesis, a novel building model reconstruction algorithm based on a dimension-reduction method is proposed, mainly including two steps:two-dimensional features organization and three-dimensional solid model generation. The advantage of this method is that it can simplify the modeling process by making each algorithm focus on individual regularization problem. In the two-dimensional features organization stage, artificial knowledge can be added easily, thereby improving the quality of building models in complex scenes.