三维激光扫描点云模型快速构建和质量评价关键技术研究

The rapid building and quality evaluation of point cloud model have been a hot topic in the field of 3D laser scanning technique. In this research, different theories and methods of the rapid building and quality evaluation for the point cloud model using 3D laser scanning will be discussed based on the theories of information entropy, finite element analysis and computer semantic identification. Firstly, the fusion algorithm for eliminating the point cloud noise of different attributes will be proposed, the feature attributes echo intensity model and characteristic of feature space fluctuation relation will be discussed, the model of key surface features of point cloud will be established. The classification methods of the linear and planar point cloud, also the ground and non-ground point cloud will studied using the combination of the weighted least squares and Hough transform model, and a fast and accurate classification algorithm would be put forward based on the fusion model of AdaBoost algorithm and decision tree. Secondly, to eliminate the influence of invalid point cloud during model building, PCA algorithm and entropy function will be used to determine the optimal neighborhood size of the point cloud, a robust normal estimation model will be established with the method of weighted local optimal projection (WLOP) using mobile least squares, and an adaptive simplification algorithm would be proposed based on the angle information entropy layering of normal vector. Thirdly, with the purpose of accelerating the efficiency of point cloud reconstruction, the theory of fine segmentation on level set point cloud based on fractal dimension and the classification method of point cloud categories on the basic of the theory of computer semantic identification would be studied resulting in proposing a rapid building algorithm of point cloud model using the GPU parallel rule based on ant colony algorithm and semantic identification. Fourthly, the error entropy model of line elements, surface elements and volume elements will be established, the constraint relational model of the geometry differences between the surface feature elements of point cloud would be derived, and a comprehensive evaluation system for the precision quantification of point cloud model and surface quality based on the error entropy will be put forward. Finally, the reliability, precision and adaptive capacity of the system of rapid building and quality evaluation of point cloud model in different application environments are going to be validated, in addition, a software system of rapid building and quality evaluation of point cloud model will be developed. The application prospects of this research would be great by generalizing the application of 3D laser scanning in the fields of cultural relic restoration, deformation monitoring and industrial product certification.

点云模型构建及质量评价是三维激光扫描技术研究热点之一。本课题首先研究点云噪声剔除融合算法；建立加权整体最小二乘与霍夫变换模型相结合的不同类型点云的分类方法；研究基于Adaboost算法和决策树融合的快速点云分类算法。第二，研究利用PCA和熵函数确定点云特征的最优邻域，建立移动整体最小二乘的加权局部最优投影的法向量估计模型，提出基于法向量夹角信息熵分层的自适应点云简化算法。第三，研究基于分形维数的水平集点云精细分割理论和计算机语义识别理论的点云类别归属方法，提出基于语义识别的GPU并行蚁群运算法则的点云模型快速构建算法。第四，研究构建点云线元、面元及体元误差熵模型，提出基于误差熵的点云模型精度量化及曲面品质的质量评价体系。最后，研究在不同应用环境下点云模型质量评价体系的可靠性，开发点云模型快速构建及质量评价系统。本课题研究成果促进三维激光扫描在文物复原、变形监测及工业产品检测等领域中的应用。

点云数据三维模型的快速重建和质量评价是三维激光扫描技术研究热点之一，受到广泛的关注。本项目针对关键技术问题展开研究，取得了一些实质性进展和成果。.本项目首先提出了自适应邻域的PCA点云法向量估计方法；针对点云表面特性易受邻域大小影响问题，提出了基于最优领域和二次曲面误差的点云非均匀简化算法，提出了基于局部形状TriLCI描述符的点云配准新方法，构建了基于连接图和配准误差达到全局最小的多视点云配准方法；提出了不同类型目标物体的点云特征提取方法。其次，在对多种经典分割方法在分割精度、效率及适用范围比较的基础上提出一种结合ＲANSAC和欧氏聚类的点云分割算法，构造了一种基于迭代高斯映射与多方法集成的室内场景点云分割技术；提出了一种基于点云数据的建筑物平面精细分割方法和自适应的地面分割算法。第三，研究了点云快速三维重建算法，构建了一种优化改进的泊松曲面重建算法，提出了基于GPU并行运算的点云数据快速三角建模方法和基于泊松函数的GPU隐式重建方法。第四，研究了地面三维激光扫描角度分辨率区别目标细节的敏感性关系，提出了角度分辨率精度评价方法，提出了利用误差熵评价光斑中点云不确定性方法。探索了三维点云深度学习网络中的解释方法，提出了可度量的量测指标用于精度评价。最后，针对应用，提出一种基于阈值判别变形和正态分布模型提取变形相结合的点云特征面变形分析算法；构建了基于三维激光扫描技术的历史建筑空间数据快速采集技术方法体系，自主开发地面三维激光扫描轮廓线自动提取及精度评定系统和点云三维建模快速构建及精度评定系统，在佛山市禅城区第一批61处历史建筑信息化管理应用，取得了良好的社会和经济效益。.项目发表论文58篇，其中，SCI 论文18篇，EI论文3篇；发明专利6个；获计算机软件著作权4个；参加国际会议1次；获测绘科技进步特等奖1项、三等奖1项，地理信息科技进步二等奖1项。