Research On Building Extraction And Model Regularization From Aerial LiDAR Data

Building identification and extraction has been the research hot pot in mapping and remote sensing fields. As an important symbol of the city landscape, buildings play a key role for the feature matching, the topographic map updating and digtal city,etc. Through the airborne LiDAR, a new earth observation technology, we can rapidly achieve the identification and extraction of3D information. The method to extract buildings automatically from massive LiDAR point clouds data has vast importance.Based on the previous researches, this study presents a point clouds filtering algorithm based on TIN model and a set of contour-based building boundary vectorization algorithm. The main contents and conclusions are as follows:(1)Spatial distribution characteristics of an object not only relates to the material properties, but also to its surface roughness In-depth analysis of the spatial distribution of the typical objects’ LiDAR point clouds gives the conclusions:The noise points mostly isolated distribute; The vegetation points distribute as cluster with large variance; The ground points distribute flat with small elevation difference; The water points are sparse because of the absorption; The buildings points distribute regularly, and the point density is uniform, while the other protrusions above the ground have fewer points. (2)LiDAR data filtering algorithm:Through the comprehensive evaluation of several classic filtering algorithms, this study propose a point clouds segmentation algorithm based on TIN model. This algorithm builds the space adjacency of point clouds according to TIN model, and segments the point clouds into homogeneous objects depending on the principle of clustering. The building objects are filtered according to the elevation variance. Three typical study areas includeing respectively terrain with slopes of about12°、irregular and complex shapes of buildings, and residential blocks, are selected in the experiments. The results show that the overall filtering accuracy for of the area with slope terrain is96.61%, and the kappa coefficient is0.9154; The overall filtering accuracy for the area with complex building is96.69%, and the kappa coefficient is0.8183; While the overall filtering accuracy for area with large blocks is93.72%, and the kappa coefficient is0.8076. The results imply that the algorithm has a good antigross error ability and has a good filtering effect for the slope terrain and complex buildings, and is suited for filtering buildings from the LiDAR data.(3)Building contour vectorization:The algorithm tracks the contour of the building footprint using Alpha-Shapes and realizes the effective extraction of inside and outside and outside contour’s of concave and complex buildings; A modified minimal direction difference based model fitting algorithm is proposed. It makes the difference of the estimated main direction of the building and those of all the sides of the initial contour polygon minmal. Experiments show that almost accuracy of the estimated main direction is69.4%, the probability of small deviations is26.5%, and the probability of large deviations is4.1%. The extraction accuracy is much higher than the commonly used method such as the Hough transform, principal component transform and statistical histogram. Besides, building contour regularization expands the polygon after preliminary regulation to make the polygon sides located at the accurate building edges.