基于特征的大场景地面Lidar点云配准

The development of terrestrial Lidar technology promotes requirement of 3D modeling and analysis based on point clouds. As the prerequisite to 3D modeling and analysis, registration for point clouds of terrestrial Lidar becomes one of the hot topics. Because ICP algorithm and its variants require meeting certain conditions such as good initial registration parameters, inclusion relation or large overlaps between adjacent point clouds, they don’t be suited for scenes of large scale. In this project, we concentrate on theory and methods for point clouds registration of large scale scenes based on features. By studying key problems such as how to decrease the number of points for Hough transform, to reduce dimensions of parameter space of Hough transform etc, we establish Hough transform methods suited for feature extraction of point clouds; We also derive models of point clouds registration based on line features and surface features and construct EIV model of registration, discussing effect of model Linearization on transformation parameters estimation. Non-linearization solving method, combining total least squares theory and robust estimation are used to improve accuracy and precision of registration transformation parameters. As to registration of multi-view point clouds, we propose new ideas and algorithms based on analytical aerial triangulation in photogrammetry theory, which will overcome distortion problem because of error spreading and accumulation .On basis of above researches and corresponding results, we design a prototype for point clouds registration and use it in registration of point clouds in digital city 3D reconstruction. Our study will provide theory basis and practical cases for registration of large scale scenes, and has important meanings for 3D modeling and analysis based on point clouds.

地面Lidar技术的发展使基于点云数据的三维重建与分析需求日益增加。作为三维建模与分析的前提，点云配准已成为研究的热点。针对常用ICP及其改进算法要求事先获取良好的配准初值，且相邻点云具有包含关系或具有大的重叠度，不适合大场景配准的缺点，本课题研究基于特征的大场景地面Lidar点云配准理论与方法。通过研究减少参与Hough变换的点数、降低参数空间维数等关键问题，建立适合点云数据特征提取的Hough变换方法；推导基于线、面特征的点云配准模型，构建点云配准EIV模型，探讨模型线性化对解算结果的影响，利用非线性化方法，将总体最小二乘与稳健估计理论应用于配准参数解算，提高参数解算的准确度与精度；借鉴摄影测量解析空三理论，提出严密的多视点云整体配准算法。基于以上研究，开发点云配准原型系统，并在数字城市三维建模点云配准实践中应用，为大场景点云配准提供理论依据与实践事例，对于三维建模与分析具有重要意义。

地面LiDAR作为一种空间信息获取技术，实现了快速获取空间立体数据的革命性飞跃，成为三维重建与分析的一个重要数据来源，需求日益增加。利用点云数据进行场景三维建模与分析，必须获得研究对象的完整数据，这通过点云配准得以实现，配准是点云数据处理中的关键步骤。. 目前点云配准主要包括以ICP算法为主的无特征配准和基于特征的点云配准。针对常用ICP及其改进算法要求事先获取良好的配准初值，且相邻点云具有包含关系或具有大的重叠度的条件限制，本课题主要研究基于特征的大场景点云配准理论与方法。课题主要开展以下内容的研究：1）点云数据特征提取；2）点云配准模型建立及配准变换参数解算；3）多视点云整体配准。通过研究，完成了基于点云数据的点特征、球体特征提取。分析平面标靶特性基础上，将模糊c-均值聚类与稳健标靶定位算法结合，完成靶心点坐标的提取；将2D Hough变换扩展至3D hough变换进行球体检测，同时针对Hough变换提取球特征存在运算量大的不足，提出“4点定位法”以快速定位球心坐标和计算半径值，并与Hough变换基本方法相结合，完成球面特征提取。从基于点的配准模型出发，推导以线的方向向量、面的法向量为基础的点云配准模型；在同时顾及点云数据系数矩阵和观测向量中存在误差的情况下，构建出点云配准的EIV模型；打破常规解算需对点云配准模型进行线性化的惯例，不再对配准模型进行线性化，从而消除线性化对参数解算精度的影响，以总体最小二乘法则为基础，提高配准参数解算精度；借鉴摄影测量空中三角测量中光束法区域网平差，提出多视点云数据整体配准算法。项目研究成果进一步丰富了点云数据处理理论与方法，可为大场景点云配准提供依据。与此同时，课题组还初步研究了基于曲面重构法的三维重建和基于参数法三维重建二种建模技术，并对二种建模技术进行对比分析，为后续点云数据应用打基础。