地面激光点云模型自动构建方法研究

发布时间：2018-04-03 06:00

本文选题：地面激光雷达　切入点：逐站配准　出处：《武汉大学》2014年博士论文

【摘要】：激光雷达(Light Detection And Ranging,LiDAR)是一种非接触的主动式的对地观测系统,它能够快速获取目标对象三维空间信息,为获取高时空分辨率地球空间数据提供了一种全新的技术手段。由多传感器同步集成的基础上构建的地面LiDAR测量系统,能够在静止状态下获取对象精细三维点云数据,广泛运用于质量检测、变形监测、文物保护等领域,已成为一种主要的空间数据获取手段。一方面,它具有数据获取速度快、点云密集的特点；另一方面,其获取的数据具有海量特性(每秒可获取上百万个点),且存在噪声和遮挡,这给地面LiDAR点云数据处理带来了巨大的挑战。目前地面LiDAR点云数据处理的主要问题在于：数据量大,冗余度高及由此产生的效率低；点云数据配准以人工配准为主,已有方法通用性差,自动化程度低。针对以上问题,本文重点研究了地面LiDAR点云数据中多测站点云数据自动配准和冗余去除方法。具体研究内容如下： (1)在详细介绍地面LiDAR基本原理的基础上,引出了地面LiDAR的数据特点和噪声问题,采用分级的方法对点云数据去噪,以降低噪声对后续工作的影响：首先根据目标对象的概略大小,去除目标外围噪声；然后根据单站点云数据的栅格结构,提出基于遮挡边界检测和空间聚类的孤立噪声点集探测方法,根据设定的阈值,滤除孤立噪声点集。 (2)介绍了点云两两配准的概念、方法分类和研究现状,指出当前基于特征的方法和ICP方法的不足,强调了标靶辅助的方法在实际工程应用中的重要作用；基于球面标靶的各向同性、易放置、精度高等优点,提出了基于遮挡边界检测和多级几何条件约束的球面标靶自动探测方法；针对探测到的球面标靶集,提出以三角形为基元、其面积夹角为相似性测度的球面标靶同名匹配策略；对于获得的同名点对,采用罗德里格矩阵解求变换参数,最终获得点云数据两两配准结果。 (3)在点云数据两两配准的基础上,介绍了多站点云数据的误差累积,给出了多站点云数据整体配准的概念、方法分类和研究现状。鉴于多站数据同时配准的高内存、计算量消耗,本文采用在逐站配准完成初始值计算的基础上,采用基于罗德里格矩阵的空间变换模型的线性化表达和同名标靶坐标相等的约束条件,构建误差方程,以间接平差的方式对累积误差进行分配,最终完成多站点云数据整体配准。 (4)针对经整体配准后多站点云数据产生的冗余问题,提出利用QMBB树对单站点云数据进行组织,利用三维格网对配准后的数据体进行虚拟划分,在局部格网单位内部,动态加载各测站所对应的数据子集,参照点到测站距离和法向因素完成冗余去除。三维虚拟格网计算将点云数据冗余的全局性问题转换为局部性问题,并能够大大节省空间格网划分所需要的海量内存空间。
[Abstract]:LiDAR (LiDAR) is a non-contact active earth observation system. It can quickly acquire 3D spatial information of the target object, which provides a new technique for obtaining high spatial and temporal resolution geospatial data.The ground LiDAR measurement system based on multi-sensor synchronous integration can obtain fine 3D point cloud data of objects in static state. It is widely used in the fields of quality detection, deformation monitoring, cultural relic protection and so on.It has become a main method of spatial data acquisition.On the one hand, it has the characteristics of fast data acquisition and point cloud density; on the other hand, it has the characteristics of massive data (millions of points per second can be obtained, and there is noise and occlusion,This brings great challenge to ground LiDAR point cloud data processing.At present, the main problems of ground LiDAR point cloud data processing are: large amount of data, high redundancy and low efficiency, point cloud data registration is dominated by manual registration, the existing methods have poor universality and low degree of automation.In view of the above problems, this paper focuses on the automatic registration and redundancy removal of multi-site cloud data from ground LiDAR point cloud data.The specific contents of the study are as follows:1) on the basis of introducing the basic principle of ground LiDAR in detail, the data characteristics and noise problems of ground LiDAR are introduced, and the point cloud data is de-noised by hierarchical method.In order to reduce the impact of noise on the subsequent work: firstly, according to the size of the target object, the noise around the target is removed; then, according to the grid structure of the cloud data of a single station,An isolated noise point set detection method based on occlusion boundary detection and spatial clustering is proposed. The set of isolated noise points is filtered according to the set threshold.This paper introduces the concept, classification and research status of point cloud registration, points out the shortcomings of the current feature-based method and ICP method, and emphasizes the important role of the target-aided method in practical engineering application.Based on the advantages of isotropy, easy placement and high precision of spherical targets, an automatic detection method for spherical targets based on occlusion boundary detection and multi-level geometric constraints is proposed.The matching strategy of spherical target with the area angle of similarity measure is the same name, and the transformation parameters are obtained by using the Rodrigo matrix solution, and the matching result of point cloud data is obtained.3) based on the pairwise registration of point cloud data, this paper introduces the error accumulation of multi-site cloud data, and gives the concept, classification and research status of the whole registration of multi-site cloud data.In view of the high memory and computational cost of multi-station data registration at the same time, this paper adopts the method of performing initial value calculation on the basis of site-by-station registration.The linear expression of the spatial transformation model based on the Rodrigo matrix and the constraint condition of the same name target coordinates are used to construct the error equation and to allocate the cumulative error by indirect adjustment.The final completion of multi-site cloud data registration.4) aiming at the redundancy of multi-station cloud data after the whole registration, this paper proposes to organize the cloud data of a single site by using QMBB tree, and to divide the registered data body into virtual parts by using three-dimensional grid, which is within the unit of local grid.The data subsets corresponding to each station are dynamically loaded, the reference point to the station and the normal factor are removed.The global problem of point cloud data redundancy is transformed into a local problem by 3D virtual grid computing, and the massive memory space needed for spatial grid partition can be greatly saved.
【学位授予单位】：武汉大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TN958.98

【参考文献】