ICP算法的改进及大规模点云配准方法的研究

发布时间：2018-05-30 20:10

本文选题：点云配准 + ICP　；参考：《中北大学》2017年硕士论文

【摘要】：三维点云配准技术是三维重建过程中的一个重要组成部分,在各个领域都有十分广泛的应用前景。比如在工业领域中,可以用它来检测物体零部件是否存在缺陷;在医疗行业中,可以用它来模拟人体器官并找出病人的病灶所在等。近些年,随着三维扫描设备的精度不断提高,要想得到物体精确的三维模型已经变得非常容易。因此,三维点云数据配准算法的研究也逐渐成为人们研究的重点。点云数据配准的过程就是把分次测量得到的不同角度、不同参考坐标系下的两个或多个点云数据通过一定的旋转和平移变换,将它们统一到相同的坐标系下,从而获得物体的完整信息并对物体进行一系列的可视化操作。目前已有的点云配准算法主要存在两方面的问题:一方面,传统ICP(Iterative Closest Points,迭代最近点)算法虽然在一定程度上能够满足人们对实验的要求,但它在选取对应点时,简单的将两个待匹配点云中欧氏距离最近的点作为对应点,这样会造成一定的错配点产生,从而影响算法配准的精度;另一方面,当点云数据的规模较大时,配准过程中会消耗大量的时间,造成配准算法实时性较差的问题。针对这些问题,本文主要从以下几点进行研究:(1)本文深入了解了传统ICP算法及其相关改进算法的配准过程及存在的一些问题,并在此基础上提出了基于旋转图像特征描述子改进的ICP算法。该算法在配准前首先对待匹配点云进行了滤波处理,在减少点云数据量的同时还保持点云的基本形状特征。然后找出两个点云的关键点,分别求出待匹配点云关键点的旋转图像特征描述子,并根据两个特征描述子的特征相似程度来确定最近点进而完成ICP配准,得到了较好的收敛效果。(2)为了有效解决点云规模较大时,配准实时性较差的问题,本文深入了解了基于GPU(Graphics Processing Unit,图形处理单元)的点云并行配准算法。详细介绍了EM-ICP算法和Softassign算法的配准过程,并结合GPU,实现了基于GPU的EM-ICP和Softassign并行配准算法,大幅度提高了点云的配准的效率,提高了算法的实时性。(3)在本文提出的改进算法的基础上设计并实现了基于改进ICP算法的点云配准系统,并通过编程的方式详细设计和分析了该系统中的每个模块。该系统主要分为点云显示、点云滤波模块与点云配准模块,其中点云配准模块使用了本文提出的改进ICP算法。
[Abstract]:3D point cloud registration technology is an important part of 3D reconstruction process and has a very wide application prospect in various fields. For example, in the industrial field, it can be used to detect the defects of object parts, and in the medical industry, it can be used to simulate human organs and find out where the patient's lesions are. In recent years, with the improvement of the accuracy of 3D scanning equipment, it has become very easy to obtain the accurate 3D model of object. Therefore, the research of three-dimensional point cloud data registration algorithm has gradually become the focus of research. The registration process of point cloud data is to unify two or more point cloud data in different reference coordinate systems into the same coordinate system by a certain rotation and translation transformation. Thus the complete information of the object is obtained and a series of visualization operations are carried out. There are two main problems in the existing point cloud registration algorithms: on the one hand, the traditional ICP(Iterative Closest points (iterative nearest points) algorithm can meet the requirements of experiments to some extent, but it selects the corresponding points. Simply taking the nearest Euclidean point in the cloud of two points to be matched as the corresponding point, this will result in a certain mismatch point, which will affect the accuracy of the algorithm registration; on the other hand, when the data of point cloud is large, Registration process will consume a lot of time, resulting in poor real-time registration algorithm. Aiming at these problems, this paper mainly studies the following points: (1) this paper deeply understand the registration process and some problems of the traditional ICP algorithm and its related improved algorithm. On this basis, an improved ICP algorithm based on rotating image feature descriptor is proposed. In this algorithm, the matching point cloud is filtered before registration, and the basic shape feature of the point cloud is kept while the data of point cloud is reduced. Then the key points of the two point clouds are found, and the rotating image feature descriptors of the key points to be matched are obtained, and the nearest points are determined according to the similarity degree of the two feature descriptors and the ICP registration is completed. In order to effectively solve the problem of poor real-time registration when the point cloud scale is large, this paper deeply understand the point cloud parallel registration algorithm based on GPU(Graphics Processing unit (graphic processing unit). The registration process of EM-ICP algorithm and Softassign algorithm is introduced in detail, and the parallel registration algorithm of EM-ICP and Softassign based on GPU is realized, which greatly improves the efficiency of point cloud registration. Based on the improved algorithm proposed in this paper, a point cloud registration system based on improved ICP algorithm is designed and implemented. Each module of the system is designed and analyzed in detail by programming. The system is mainly divided into point cloud display, point cloud filtering module and point cloud registration module, where the point cloud registration module uses the improved ICP algorithm proposed in this paper.
【学位授予单位】：中北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】