【作者】 刘科；

【导师】 马洪超；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2021， 博士

【摘要】 建筑物作为城市中最重要的组成元素之一,其三维结构信息以及空间位置信息一直是重要的地理信息数据。随着当前智慧城市快速发展,实际生产与应用中对高效、自动化的建筑物地理信息数据获取提出了越来越高的要求。机载激光雷达(Light Detection And Ranging,LiDAR)是一种主动式航空遥感对地观测系统,其可以在短时间内获取海量高精度、高密度的地物表面三维点云数据,以机载LiDAR点云为数据源可以提取更多建筑物相关信息。而LiDAR点云数据散乱不规则,缺乏颜色、纹理、光谱等信息,同时复杂的城市场景以及多样的建筑物结构,给点云数据处理带来了难度,在一定程度上阻碍了点云数据在实际工程中的应用。因此,本文以机载LiDAR点云为数据源,基于图割理论研究建筑物分割与轮廓线提取,解决建筑物提取、屋顶面分割与轮廓线提取过程中的技术难点。完成的主要工作内容如下:1.提出一种结合图割与后处理的建筑物提取方法。该方法先对原始点云进行滤波处理得到地面点与非地面点,再使用非地面点及点云特征构建无向权重图,根据最小割原理对无向权重图进行分割粗提取建筑物。在后处理中,使用高程约束、限制区域增长、最大夹角约束与一致性约束精化粗提取结果。选取具有不同密度、不同区域的机载LiDAR点云数据进行实验,证实了本文方法的可行性,为后续建筑物各类应用提供可靠数据。2.提出一种基于图割优化的单体建筑物分离方法。首先使用三维空间欧氏距离聚类方法对建筑物点云数据进行聚类,得到初始单体建筑物;再构造一个考虑点到初始聚类中心距离、近邻点连续的能量函数;最后使用图割算法对能量函数进行优化,实现单体建筑物分离。采用建筑物分布不同的测区数据进行实验,实验结果表明本文方法可以有效克服因点云密度不均导致建筑物分离失败的问题。3.提出一种结合迭代区域增长与改进全局能量优化的方法进行建筑物屋顶面分割。首先使用连通域分析法对建筑物点云数据进行处理得到单体建筑物,再使用迭代区域增长算法对单体建筑物进行屋顶面分割,得到初始屋顶面分割结果。再使用改进全局能量优化算法优化屋顶面分割结果,在改进的能量优化算法中,对每个点只需计算其到特定屋顶面的代价项,减少计算量。使用具有不同点云密度、屋顶结构的测区数据进行测试,实验结果表明本文方法可以准确分割建筑物屋顶面。4.提出一种基于主方向软约束的方法提取建筑物轮廓线。首先使用边界跟踪Alpha shapes算法快速提取得到有序轮廓点,再使用迭代区域增长与聚合层次聚类方法获取初始轮廓线,最后使用建筑物主方向对初始轮廓线进行规则化。使用不同测区的数据进行建筑物轮廓线提取实验,实验结果表明本方法可以准确提取不同类型建筑物轮廓线,具有较强的实用性。更多还原

【Abstract】 As one of the most important constituent elements in the modern city,threedimensional(3D)structure information and spatial location information of buildings have always been important geographic information data.With the rapid development of smart cities,efficient and automatic acquisition of building geographic information is required both in practical production and various applications.Airborne Light Detection And Ranging(LiDAR)is an active aerial remote sensing earth observation system,which acquires a large amount of high-precision and high-density 3D point cloud data of objects surfaces in a short time.Therefore,more building-related information can be extracted from LiDAR point cloud data.LiDAR data are scattered and irregular,lack color,texture,and spectral information.As a result,it is a challenging task to process point cloud data and apply point data in practical applications due to the above problems and complex urban scenes,which hinders the application of point cloud data in practical engineering.Thus,this paper takes airborne LiDAR point cloud as the data source and studies building segmentation and contour extraction based on graph cuts algorithm to solve the technical difficulties in the process of building extraction,roof plane segmentation,and contour extraction.The main works accomplished are as follows.1.Building extraction based on graph cuts and post-processing is proposed.First,non-ground points and ground points are obtained from raw LiDAR data after filtering.Then an undirected weighted graph is constructed using non-ground points and point features,and initial building points are obtained by cutting the constructed undirected weighted graph based on min-cut.In the post-processing,restricted region growing,constraints of height,the maximum intersection angle,and consistency are used to refine initial building extraction results.LiDAR data with different densities and regions are selected in the experiments.Experimental results confirm the feasibility of the proposed method and reliable data can be provided for various applications of subsequent buildings.2.Individual building segmentation based on graph cuts is proposed.3D Euclidean distance clustering technique is used to obtain initial individual buildings.Then,an energy function considering the distance of the point to the initial clustering center and the connectivity between adjacent points is constructed.Last,initial individual building segmentation results are refined by minimizing the constructed energy function via graph cuts.Data of areas including building distribution is used for experiments,and experimental results show that the proposed method can accurately obtain accurate individual buildings and is insensitive to uneven point cloud density.3.Building roof segmentation combining iterative region growing and improved global energy optimization is proposed.First,individual buildings are obtained via the connected component analysis method.Then,an iterative region growing is used to obtain initial roof planes.After that,initial roof planes segmentation results are optimized via an improved global energy optimization.In the process of optimization,the data cost between each point and only specific roof planes are calculated to decrease computation cost.Experimental results of different point cloud density and building structures show that the proposed method can accurately segment roof planes.4.A method based on the soft constraint of building dominant direction is proposed to extract building contour.First,boundary tracing Alpha shapes is used to extract ordered boundary points efficiently.After that,the initial contour is obtained via iterative region growing and agglomerative hierarchical clustering.Finally,the initial contour is regularized under the soft constraint of building dominant direction,and regularized building contour is generated.Experimental results show that the proposed method can extract different types of building contour with high accuracy.更多还原