Object Oriented Classification And Building Extraction Using Airborne LiDAR Point Clouds

In the airborne LiDAR point cloud analysis, it is an important prerequisite for automatic three-dimensional reconstruction to accurately detect the point cloud of the buildings. The loss of information from a point cloud to a digital surface model (DSM) through rasterization is much and this processing is also irreversible. Further classification of point cloud is required even if the building area is detected. Therefore, the direct analysis of the point cloud can effectively achieve automatic three-dimensional reconstruction, but point-based classification method with local information will result in error classification caused by salt-and-pepper noise. Therefore, this paper uses object-oriented classification for building extraction. Considering about building extraction, one of the most difficult part is the distinction between buildings and vegetation, especially when they are adjacent. In order to solve this problem, on the basis of the object-oriented point cloud analysis, we use the structural characteristics of the building to classify it under the optimization framework.This paper mainly researches the building extraction based on airborne LiDAR point cloud. The main contents of this paper are as follows:(1) A fast and accurate point cloud segmentation method for LiDAR point cloud is considered, and a fast plane segmentation algorithm based on fusion profile analysis, model fitting and region growing is proposed:Cross Line Element Growth (CLEG). The Meanshift clustering is performed for the nonplanar points, and then the result of the segmentation is obtained, so that each point belongs to an object.(2) Feature extraction method is researched. In this paper, principal component analysis (PCA) is used to do the object planarity analysis. And the roughness of the object is obtained by Douglas line segmentation in profile analysis. Z-Maximally Stable Extremal Region (Z-MSER) is used to analyze whether the region is stable, by considering the elevation direction slice to get the Z-direction region change. According to the clustering of the objects, the wall points are extracted. And the contour points are obtained according to the points of elevation transition, which is used for further contour line extraction. In order to obtain stable result, the saliency map of the line is obtained by tensor voting, and then the line is extracted on the saliency map by ELSD. Finally, the multi-label optimization based on Graph Cuts is used to get the result of the final contour extraction.(3) A coarse to fine extraction strategy is proposed in this paper to extract buildings from LiDAR point cloud. In the segmentation step, the result of plane extraction is obtained by plane segmentation, then the non-planar points are clustered to get the objects. In the extraction process, the plane object is divided into two categories, namely, building facings and non-building facings (most of which are ground, and the rest are small surface structure of the ground). In this paper, we extract flaky area using regional growth algorithm in the first place. The larger area is the ground, and then the Triangulated Irregular Network (TIN) is used to get the initial ground, and the results are obtained by Graphcuts iterative optimization. The plane of these buildings is the initial value of the next extraction. Three types of features are calculated for each non-planar object based on the initial value, including:(a) height to the ground; (b) planarity, linear features, etc; (c) roughness. The result of Douglas line segmentation is used to calculate the probability of belonging to the building. After constructing the neighbor relation of the object, the extracted contour line, the wall line and the Z-MSER are taken as constraints. In the framework of the Graphcut, the result is classified by Min cut/Max flow.(4) The visual information of the point cloud is merged to study classification methods. The method based on visual reconstruction can be regarded as the method of space classifying. Based on this, a unified framework for building extraction and reconstruction is proposed. Under the framework of object-oriented classification, the partitioned space is considered as the optimized unit, and the result of spatial classification is obtained, in other words, result of point cloud classification is obtained. The specific process is as follows:obtain the border lines of initial area based on the initial building area. Next, extract contour lines. Regularize the lines according to the extracted principal direction. And then, do space partitions using Binary Space Partition (BSP). After calculating the space where the point cloud is located, the weight of each space is calculated according to the object. Calculate the visibility weights based on building points and ground points. After the smoothing weight is calculated, Min cut/Max flow is used to get the spatial classification result, and then the classification result is obtained. The surfaces of buildings should be obtained according to the intersection of patches of space.