【作者】 张志超；

【导师】 李德仁；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2010， 博士

【摘要】 目前,众多应用领域对城市建筑物三维模型都有较大的需求。但,目前城市建筑物的获取效率却较低,为了提高城市建筑物三维重建的效率,本文对现有的建筑物重建方法进行总结,并研究融合目前较新型的机载与地面LIDAR数据进行建筑物屋顶面以及立面的三维重建,获取较为精细的建筑物三维模型的方法。本文在总结和分析了以往从机载以及地面LIDAR数据中进行城市地区建筑物三维重建的相关技术和理论的基础上,掌握该研究方向的现有方法的问题与不足,然后针对LIDAR数据重建建筑物模型的特点,利用点云分割和形状/分裂语法为重建的重要思路,建立了一套融合机载与地面LIDAR数据进行城市地区建筑物三维重建的理论体系和处理流程,提出了高精细建筑物三维重建的解决方案,最后通过实验验证了本文解决方案的有效性和可行性。本文的主要研究如下：(1)研究了地面LIDAR设备的工作原理,并基于该原理对地面LIDAR数据进行扫描线重构,进而根据扫描线恢复了LIDAR数据中点的拓扑关系。并根据该拓扑关系生成了单站地面LIDAR数据的点云影像。基于点云影像,可以借鉴大量已有的成熟图像处理算法对LIDAR数据进行处理,本文基于点云影像提出了一系列的数据处理方法,如法向量计算、点云数据分割等,并通过实验验证了该方法的有效性和高效率。(2)采用机器学习的方法对机载LIDAR数据进行建筑物检测,利用点云数据的局部微分几何特性来度量其属于建筑物顶面数据的可能性,并结合支撑向量机方法来对屋顶面数据的微分几何特性系数进行训练,从而得到建筑物顶面数据检测的模型。为了从地面LIDAR数据中检测出建筑物数据,将最小二乘平面拟合的残差作为LIDAR数据中点的平坦度度量,区分开了LIDAR数据中的规则分布数据与杂乱数据。然后利用法向量平滑度约束方法对点云数据进行分割,可以将点云数据分割成多个规则点集,再通过点集的多种属性约束从规则点集中识别出建筑物数据。(3)结合现有的地理空间定位资源,提出了一种半自动的机载与地面LIDAR数据配准方法。在交互式的获取了地面LIDAR数据扫描站点在机载数据中的概略位置后,通过两步法来分别解求两种数据配准的旋转和平移参数,先分别在两种数据的概略对应位置选取一定范围内的点云进行平面拟合,通过两个对应的平面来确定平移参数,然后在地面扫描站点附近自动搜索机载LIDAR数据中的屋顶轮廓空洞,以及地面LIDAR数据中的立面,利用从空洞轮廓中提取的机载建筑物立面与地面LIDAR数据中的建筑物立面进行匹配,得到配准的旋转参数,完成机载与地面LIDAR数据的配准,并通过平面之间的夹角对配准的精度进行了分析。(4)采用相邻点法向量相似聚类的思路对屋顶数据进行面片分割,同时采用规则格网最邻近点方法提取面片的轮廓线,对于平顶型的屋顶,直接利用数据驱动的方法,对轮廓线进行基于直线投票的规则化以及合并来生成屋顶模型,对于存在屋脊或者金字塔形的屋顶,通过面片的拓扑图以及面片与水平面的夹角来生成形状语法,通过形状语法生成屋顶面模型。(5)定义了建筑物立面多种要素的多种独有属性以及相互关系,利用这些规则关系约束对分割出的面片进行类型识别,识别出立面的多个要素。并利用规则格网法对建筑物立面的窗户进行提取,得到窗户的大小,位置以及类型。从门窗的分布情况提取出建筑物立面的结构信息,基于结构信息自动构建形状语法,再利用形状语法编译器生成建筑物立面的模型。(6)融合机载LIDAR数据中提取的屋顶高度以及轮廓线信息对建筑物部分缺失立面进行推理,同时结合地面LIDAR数据提取出的较精细立面对建筑物轮廓线进行精化。以达到融合两种数据构建更优建筑物模型的目的。(7)结合两份具有一定代表性的机载与地面LIDAR数据对本文提出的方法和理论进行了实验,从中提取出了具有较精细立面的建筑物模型,并对建筑物轮廓线进行了精化,验证了本文所提出的理论和方法的可行性和有效性。本文并没有利用影像相关的信息,仅仅依赖从LIDAR数据中提取的几何信息,再结合自上而下的思路来重建建筑物,该方法为大规模快速更新提供了新的思路。目前的研究成果尽管几何信息比较丰富,但是由于没有映射纹理影像,其仿真程度仍有较大的提升空间。更多还原

【Abstract】 There are increasing demands in 3D models of City buildings in many fields. However, the efficiency of 3D building modeling is not high enough to catch up with the huge demands. This paper present a integrated method which fuse the newly emerging airborne and terrestrial LIDAR data. By this method, we combine the extracted roof and the facade of the building for a better 3D model of buildings.This paper summarizes and analyzes the past related work. Based on the the overview, the problems and shortcomings of existing methods are mastered. Then based on the special property of LIDAR data, we use a point cloud segmentation and shape/split grammar as the key ideas for the reconstruction of city buildings. Also, we proposed a set of fusion method for airborne and terrestrial LIDAR data to reconstruct the buildings in urban area. The proposed construction theory is aimed to build high detailed building facade. Finally, we push on a experiment to make the avalibility of our process and method. The main research work are summarized as follows:(1) Terrestrial LIDAR equipment working principal is deeple learned, and the scan line reconstruction is carried out based on the working principle, and then the topology of points is reconstructed under the scanning line. A point cloud image is generated according to the topology of points in a single-station LIDAR data, Based on point cloud image, we can draw a large number of existing mature image processing algorithms for processing LIDAR data, this image based on a series of point cloud data processing methods, such as normal vector calculation, point cloud segmentation, and verified by experiments the method is effective and efficient.(2) The building detecting from airborne LIDAR data is done by a machine learning methods. Using the local differential geometric properties to measure the top of their data is the possibility of building, combined with support vector machines to the roof surface coefficient differential geometric properties of data for training, which are building top of the data detection model. LIDAR data from the ground to detect the building data, the least squares plane fitting residuals as the LIDAR data, the flatness of the mid-point measure, to distinguish the rules of distribution of LIDAR data in the data and unstructured data. Then the normal vector smoothness constraint method to segment the point cloud, point cloud data can be split into multiple rules set, and then through the restriction point set of a variety of attributes identified from the rules for building data point set.(3) combined with existing geo-spatial positioning of resources, a semi-automatic airborne and ground LIDAR data registration method. Access to the ground in the interactive data scanning LIDAR site in airborne data in the approximate location, respectively, by two-step solution to seek two kinds of registration is rotation and translation parameters, the first compendium of data at the corresponding position two Select a range of point cloud fitting plane through the two corresponding plane to determine the translation parameters, and then scanning the ground near the site automatically searches for airborne LIDAR data in the roof of an empty profile, and ground LIDAR data in the facade, Use empty profile extracted from the airborne and ground LIDAR building facade building facade data matching, the rotation parameters obtained registration to complete airborne and ground LIDAR data registration, and by plane between angle on the registration accuracy was analyzed.(4) using neighbor method similar to clustering vector data, the idea of the roof surface film division, while using regular grid point nearest to extract the contour patches, the flat top type roof, the direct use of data-driven method, the contour lines to vote on the rules of linear and combined to generate the roof model, the existence of pyramid-shaped roof, the roof or through the surface film of the topology and the surface film and the angle between the horizontal plane to generate the shape grammar, by the shape grammar roof surface model generation.(5) defines the various elements of the building facade as well as the many unique properties of the relationship between the use of these rules binding relations between the surface of the segmented type of film for identification, identified a number of facade elements. And use Rules grid method the windows of the building facade was extracted by the window size, location and type. From the distribution of doors and windows to extract structural information of the building facades, based on structural information automatically build the shape grammar, syntax and then use the compiler to generate the shape of the building facade model.(6) integration of airborne LIDAR data to extract the roof height and contour information on the missing portion of a building facade reasoning, combined with terrestrial LIDAR data to extract more detailed outline for the building up in the face refinement. To achieve better integration of two kinds of data to construct the purpose of building models. (7) combined with two representative of airborne and terrestrial LIDAR data, the proposed method and the theory was, to extract out of a more sophisticated model of the building facade, and the footprint of the buildings are refined, verified that the the feasibility and effectiveness of the proposed theory.This paper did’t use any optical images, but reply only on the LIDAR data to extract geometric information,combining with a top down way to do the reconstruction of buildings. The proposed method approached a new thought about updating the large scale 3D urban scene. Now the result with detailed geometric information has a great simulating space for that no texturing mapping is done yet.更多还原