A Non-Parametric Method for 3D Building Reconstruction Using Airborne Laser Scanner Point Clouds

In recent years, the techniques of surface representation have been changed and three-dimensional models have been replaced with two-dimensional maps. Airborne laser scanner as a powerful system has been known in remote sensing as a valuable source for 3D data acquisition from the Earth’s surface which can mainly be employed for 3D reconstruction and modeling. 3D reconstruction of buildings as an important element of 3D city models, based on LiDAR point clouds has been considered in this study. A new non-parametric method is proposed for generation of 3D model of buildings with flat and tilted roof. The approach comprises of four steps for 3D building reconstruction as: (A) Simultaneous building extraction and segmentation, (B) Edge detection, (C) Line approximation, and (D) 3D modeling. In step (A) a multi-agent method is proposed for extraction of buildings from LiDAR point clouds and segmentation of roof points at the same time. In this method five criteria such as height values, number of returned pulses, length, normal vector direction based on Constrained Delaunay Triangulation, and area are utilized. Next, in step (B) the edge points of roof segments are detected. Points of triangles having no neighboring triangles are extracted as primary edge points. In the extraction process, noises, external objects, and tree points on the roofs are eliminated. It is an advantage of the proposed method, however it leads to create the undesired edge points. There is the same problem regarding to segments which contain overlap with each other (like flat building). These undesired edge points as internal points are known and must be removed. In this paper, a method named “Grid Erosion” is employed for removing these internal points and therefore finding real edge points. After detecting the final edge points, a RANSAC-based algorithm is employed to approximate building lines in step (C). RANSAC is a powerful technique in line fitting and in comparison with general least square method, especially with noisy data, it provides robust results. In order to reduce the sensitivity of RANSAC to select parameters and no need for heavy post-processing, edge points are grouped by considering the angle between two consecutive connecting convex points. After classification of edge points, a RANSAC algorithm is separately applied on each classified edge-points group to produce primary lines. The regularization constraints should be applied on primary lines to generate the final lines. Finally, by modeling of the roofs and walls, 3D buildings model is reconstructed in step (D). The proposed method has been applied on the LiDAR data over the Vaihingen city, Germany. Building roof model is manually digitized from LiDAR point clouds and compared with building roof models that reconstructed using proposed method. In model reconstruction, the dominant errors are close to 30 cm which is calculated in horizontal distance. The main advantage of this method is its capability for segmentation and reconstruction of flat buildings containing parallel roof structures even with very small height differences (e.g. 10 cm). The results of both visual and quantitative assessments indicate that the proposed method could successfully extract the buildings from LiDAR data and generate the building models.