地面三维激光扫描点云场景重建方法研究

发布时间：2018-03-28 11:27

本文选题：三维点云　切入点：点云分割　出处：《中国地质大学》2017年博士论文

【摘要】：随着地理空间信息服务产业的快速发展,对地理空间数据产品的要求也越来越高。地理空间数据正朝着大信息量、高精度、可视化和可挖掘的方向发展。三维激光扫描作为一门新兴的测绘技术,是测绘领域继GPS技术之后的又一次技术革命。该技术又称实景复制技术,能在短时间内对感兴趣的区域建立详尽的三维模型,并进行定量分析,为广大测绘领域的科研工作者和工程技术人员提供了解决问题的新手段。目前国内三维激光扫描技术正逐渐应用于文物保护、工程计算、地形测绘、虚拟现实、变形监测等领域。然而,在运用三维激光点云进行三维场景重建时还存在着许多问题,例如,在地形绘制场景中,研究者更关注点云的绘制效率,而现有的三维点云重建方法虽然提出了点云精简算法,但很少有人关注点云所代表的某类地物特性,缺少对点云数据进行分类精简和重建的方法;在文物场景重建过程中,研究者更关注对缺失数据部分的重建精度,而现有方法的预测方法仅采用了支持向量机模型对缺失数据进行预测,对缺失数据部分的重建精度较低;在变形监测中,研究者更关注不同场景之间的差异化描述,而现有方法缺少对变化场景的整体描述的方法,难以精确描述不同场景之间相同区域的的整体差异。为此,本文针对三维激光扫描场景重建过程中存在的问题进行深入研究和探讨,所做的主要研究工作及取得的研究成果如下:1.在进行多幅场景数据重建时,需要对不同场景的点云数据进行配准,在系统地分析现有基于点云几何特征的点云配准方法基础上,进行了基于点云特征点的整体点云初配准,该方法通过对特征点SIFT的提取,首先得到一组匹配点对,然后运用SVD矩阵分解算法求出初始旋转矩阵R和初始平移矩阵T。在此基础上对现有的ICP精配准算法进行改进,针对ICP算法在高斯-牛顿近似值误差较大时收敛速度慢的问题,采用LM算法对其投影矩阵进行优化,提出了基于SIFT特征点的LM-ICP点云配准算法。此外,在实际配准过程中,由于多幅场景的比例尺不尽相同,针对现有的ICP算法不能进行多尺度点云配准的问题,提出了基于SIFT特征点的多尺度点云场景配准算法。通过实验验证了算法的有效性,取得了良好的配准效果。2.在进行单幅点云场景数据重建时,或者当多幅点云场景数据配准成为单幅点云场景时,由于三维激光扫描仪扫描的点云数据通常都是海量的三维数据,如何对点云数据进行简化将影响点云精简的效率以及最终的点云重建效果。本文提出了一种点云数据精简改进算法,根据扫描对象外在属性(如变形区域和非变形区域,重点描述区域或非重点描述区域)来对三维点云数据进行分类,将非重点区域的点云数据采用包围盒精简,对非重点区域的点云数据则不进行数据精简。最后,将精简后的点云和未精简的点云采用三角网逐点插入法进行三维场景重建。实验结果表明该方法三维点云场景的快速重建提供了新的解决思路和方法。3.在进行变化点云场景重建时,需要采用合理的方法提取变化场景的点云数据和描述变化场景的点云数据模型。现有的基于GPS采集仪的变形观测由于受到观测点数的影响,不能对变化区域进行整体描述,因此,本文提出并分别采用区域增长法和形态学滤波算法提取点云的地面部分数据点,将不同变化区域的场景数据,采用数学模型进行描述,利用不同场景数据之间的相互夹角变化来描述场景中相应面数据面的整体变化情况。此外,由于现有的改进形态学滤波算法不适用于起伏变化的地面数据提取,因此提出了改进的形态学点云滤波算法,为三维场景变化描述提供了新的解决思路。4.在进行缺失场景的点云数据重建时,由于仪器本身不足或人为干扰,容易出现点云数据丢失的情况。因此,在进行建模之前,需要对点云缺失部分数据进行修补。由于不同缺失部分的点云数据特征不同,采用何种方式正确对其进行填补至关重要的问,本文选用了基于粒子群优化的SVM模型和遗传算法优化的SVM模型对其缺失部分的点云数据进行预测。实验结果表明,该方法能正确预测点云数据,为缺失点云数据的生成提供了新的有效方法。
[Abstract]:With the rapid development of information service industry in the geographic space of geospatial data products are increasingly high requirements. Geospatial data is facing the large amount of information, high accuracy, visualization and mining direction. A new survey technology for 3D laser scanning is a technology revolution after GPS the field of Surveying and mapping. The technology is also called real copy technology, can be established for the region of interest of 3D model detailed in a short period of time, and quantitative analysis, provides a new solution for the majority of the field of Surveying and mapping scientific researchers and engineers. At present, the 3D laser scanning technology is gradually applied to the protection of cultural relics and the engineering calculation, topographic mapping, virtual reality, the field of deformation monitoring. However, 3D scene reconstruction using 3D laser point cloud when there are still many problems, for example, in the terrain Draw the scene, researchers pay more attention to the rendering efficiency of point cloud, and the 3D point cloud reconstruction method although the existing point cloud simplification algorithm, but few people pay attention to features some characteristics of point cloud is represented by the lack of classification and reconstruction of streamlining point cloud data; in the process of cultural reconstruction, researchers pay more attention to the lack of data part of the reconstruction accuracy, and the prediction method of existing methods only using a support vector machine model to predict the missing data, the lower part of the missing data on the reconstruction accuracy; in deformation monitoring, the researchers pay more attention to the differences between different scenes of the description, and the existing methods lack method the overall change of scene description, it is difficult to accurately describe the same region of the overall differences between different scenes. Therefore, in view of the existing 3D laser scanning scene reconstruction problem in the process of in-depth The research and discussion, the main research work and achievements are as follows: 1. in the multi scene data reconstruction, the need for registration of different scenes of point cloud data, the systematic analysis of the existing cloud point cloud registration method based on geometric features based on points, in the overall point cloud feature the initial point cloud registration based on this method, the feature extraction of SIFT, first get a set of matching points, then use the SVD matrix to calculate the initial rotation matrix R and the initial translation matrix T. based on the ICP precise registration algorithm, improved the existing decomposition algorithm, ICP algorithm for Gauss Newton approximation the value of the slow convergence speed of large error, LM algorithm is used to optimize the projection matrix, the SIFT feature points LM-ICP algorithm based on point cloud registration. In addition, in the actual registration process, due to multiple scene proportion Size is not the same, the multi-scale ICP algorithm for point cloud registration not existing problems, put forward some multi-scale SIFT feature point cloud registration algorithm based on scene. Through the experiments verify the effectiveness of the algorithm, and achieved good effect of registration.2. in the single point cloud reconstruction of scene data, or when multiple point cloud registration has a single scene data point cloud scene, the 3D laser scanning point cloud data are usually 3D data, how to simplify the point cloud simplification will affect the efficiency and the effect of point cloud reconstruction of point cloud data. This paper presents an improved algorithm of point cloud data reduction. According to the scanning object external attributes (such as deformation zone and non deformation area, focusing on regional or non key zone) to classify the 3D point cloud data, the point cloud data in key areas with the package Box streamlined for non key areas of point cloud data is data reduction. Finally, the point cloud streamline the point cloud and not streamlined with triangular net point insertion method for 3D scene reconstruction. Experimental results show that the method of rapid reconstruction of three-dimensional point cloud scene provides new ideas and methods to solve.3. the change of point cloud reconstruction, point cloud data model to extract the changing scene using reasonable methods of point cloud data and describe the change of the scene. Based on the existing deformation observation GPS acquisition instrument due to the impact of observation points, not overall description of the change area therefore, this paper proposes and ground data points using region growing method and morphological filtering algorithm to extract the point cloud, the scene data of different regional variations, a mathematical model is described, using the mutual data between different scenes The change in angle to describe the change of the whole scene in the corresponding data. In addition, because the ground data improved morphological filtering algorithm is not suitable for the existing fluctuating extraction, therefore proposed morphological point cloud filtering algorithm, described provides a new solution for.4. in the absence of scene reconstruction for point cloud data the 3D scene changes, because the instrument itself is insufficient or prone to jamming, point cloud data loss. Therefore, in the modeling before the need to repair the missing part of point cloud data. Because of the different part of the point cloud data characteristics are different, the manner in which the correct fill in this paper is asked. The prediction of particle swarm optimization SVM model and genetic algorithm to optimize the SVM model for the missing part of the point cloud data. Based on the experimental results show that this method is able to The prediction of point cloud data provides a new and effective method for the generation of missing point cloud data.

【学位授予单位】：中国地质大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：P225.2

【相似文献】