The Quickly Automated Extraction Of Building Facades Information From Mobile LiDAR Point Clouds

Mobile Light Detection and Ranging (LiDAR) has been one of the most important applications for extracting building facade information for digital city modeling due to its high efficiency, high precision, high density and real-time obtaining 3D spatial coordinates. The digital city point clouds acquired by mobile LiDAR are massive, high density and three-dimensionally discrete. The objects to be measured include ground points, poles, low shrubs, pedestrians, building facades, low buildings and gross errors etc. It becomes one of the important missions for digital city construction to quickly automated extract building facades information from mobile LiDAR point clouds, which is also a troublesome problem for the post-processing of mobile LiDAR point data.As to the extraction of the building facade information from mobile LiDAR point clouds, the data index should firstly be constructed, followed by the automatic filter of the ground and non-ground points. Then comes the classification of the building point clouds from unground points. Finally, the building facades information is extracted. But it is of low efficiency for data index constructing with normal Octree or 3D grid methods because the mobile LiDAR point clouds are mass data and three-dimensional discrete. And the ground and non-ground points are difficult to be filtered correctly in the zone of low hills and shrubs since the threshold of scan-line or fixed slope method can’t adjust with terrain automatically. For the filtered non-ground points, it is difficult to classify building points correctly and automatically from complicated points which are concerned with buildings, trees, poles, pedestrians and etc. In the process of the extracting of building facades with normal Random Sample Consensus (RANSAC) algorithm, partial planes are prone to being over-segmented, and some pseudo-planes are probably produced.With regard to the quickly automated extraction of building facades information, the research takes advantage of the hash table to optimize the storage of Octree point data with the introduction of 3D grid and linear Octree, and adopts the 3DGrid_Hash_Octree algorithm to construct high-efficient index of mobile LiDAR data to realize the massive index of discrete 3D points. In the process of ground and non-ground points filtering, the terrain adaptive filtering algorithm (TAFA) for mobile LiDAR data is proposed after the improvement of the seeds selecting and region growing rules according to the continuity or local smoothness of the city ground. Thus, the problem is solved that the threshold of scan-line and fixed slope methods can’t adjust with terrain, and TAFA improves the accuracy of automatic classification for ground and non-ground points. After the classification of the non-ground points, the knowlegde base of classification and distinguish rules will be built according to the objects’ natural feature, spatial feature and topological relationship, with the combination of such common features as point density, spatial distribution and spatial form of similar objects. In this way, buildings, trees, street lamps and pedestrians will be recognized and classified automatically and the problem of low efficiency of human assistance is solved. Finally, by analyzing the reasons of over-segmentation, coplanar points and pseudo-planes, the improved-RANSAC (I-RANSAC) algorithm is proposed. Based on I-RANSAC, the over-segmented planes can be merged with the help of the plane normal and distance, and the pseudo-planes can be removed by fixed distance neighbor (FDN) ball, and the planes information can be extracted automatically from building facade point clouds.This research develops a new software concerning the post-processing of mobile LiDAR point clouds, for which a series of experiments have be conducted including the fast construction of points data index, the filtering of ground and non-ground points, the automatic classification of ground objects and the automatic extraction of building facade information. All the data used in those experiments is either acquired by the Trimble mobile LiDAR or the necessary and reliable analog data. The experimental results show that the efficiency of index constructing and searching by 3DGrid_Hash_Octree is enhanced by 35% and 55% than 3DGrid_Octree algorithm. Consequently, the problem of fast index and searching of massive and discrete 3D point clouds is solved. The terrain adapted filtering algorithm realizes the automatic and correct filtering of ground and non-ground points in complex terrain regions. It can provide precise feature point sets for objects classification, such as building, trees and etc. And the precise ground point clouds are suitable for road extraction. The knowledge-based automatic objects classification method can automatically and successfully recognize and classify different objects, such as buildings, trees, street lamps and pedestrians. The correct building facade extraction and segmentation are realized by I-RANSAC method, which can provide correct and precise point models for accurate plane boundary extraction and digital 3D modeling. The experimental data shows that the algorithm of Quickly Automated Extraction of Building Facades Information (QAE_BFI) for mobile LiDAR point clouds is correct and effective, and the software based on QAE_BFI is feasible and practical.