Research On Plane Feature Extracting Technology Of The Buildings Point Cloud From Laser Scanning

Laser scanning technology which can capture high resolution spatial data is graduallybecoming a new research hotspot in the “Digital City” and3D reconstruction of buildings.However, the data acquired by laser scanning technology is big, which brings huge challenge forcomputer processing. At present, the efficient and accurate methods of reconstruction is in theexploration state for the vast of point cloud data. So, this paper is focused on automaticsegmenting buildings and extracting planes of building fa ade at the background of3Dreconstruction of buildings. The details of the paper are as follows:1. For the linked problem of objects point cloud, this dissertation propose a non-groundpoints segmenting algorithm of basing on cylinder neighborhood, separating ground points(including low objects on the ground) with non-ground points. According to the distributionfeature of the ground points and objects point, this method proposes two steps. The first is tocalculate the height difference of the points in the cylinder neighborhood which is compared withthe fixed threshold value to initial segmentation between ground points and non-ground points.Secondly, initial segmentation is easy to mistake partly non-ground points (e.g., scattered leaves)as ground points, so this method utilizes Euler distance clustering algorithm to the achievingground points by the initial segmentation for getting real ground points, and merge the rest pointsand non-ground points which is achieved by the initial segmentation to the non-ground points.Experiments confirm the feasibility of this algorithm, and analyze the influence to the results ofsegmentation about the cylinder radius for the segmentation algorithm;2. For the problem of incompletely segmenting buildings point cloud by the algorithmbased on image, we propose the building point cloud segmentation algorithm based on thehistogram of target curvature. Firstly, we use Euler distance clustering algorithm to cluster thenon-ground points, and then we cluster the points which meet the certain spatial distance to thesame class and achieve objects clustering. Secondly, the targets with small volume are excludedby calculating the volume of each class. Finally, Since the remaining goals also containsnon-building, such as groups of trees, we count the curvature histogram of the remainder targetsand segment the building points by the largest probability of zero in the curvature histogram.Experimental results show that the method can effectively solve the incomplete problem ofbuilding point cloud segmentation;3. A plane fitting algorithm based on Accelerating hypothesis generation for multi-structurealgorithm is introduced to extract the building plane. Firstly, the proposed method uses randomsampling to generate a set of models. For each model, all pixels have been sorted by calculating their residuals. As a consequence, the sampling process is guided by the inliers conditionalprobability distribution, which is calculated by residual sorts of each model. Experimental resultsshow that the method can accurately extract planes in the point cloud data, and it can obtainhigher efficiency of sampling than the RANSAC, especially in the point cloud of multi-structure.