论文总字数：22217字

摘 要

随着现代化工业需求的不断提升，机器视觉成为当下的研究热点。点云配准作为其中重要的一环，为工业机器人的工件定位提供了方法，如何快速且准确的实现工件定位成为点云匹配中的重点和难点。本文基于Kinect对工件定位方法进行研究，主要工作如下：

（1）通过Kinect获取RGBD数据，获得工件点云。对点云数据进行预处理：利用直通、体素、统计三种滤波方式大幅降低点云数量；采用欧式聚类分割算法对点云数据进行分割，得到单个工件点云子集。

（2）对估计工件点云进行法向量，针对K邻域确定问题，采用Kd-tree空间索引方法加速法向量估计。针对PFH特征描述子计算量大的问题，提出采用FPFH特征描述子对对应点对进行特征描述，提高计算效率。

（3）分析点云匹配原理，针对ICP算法初值要求高、易陷入局部最优的缺陷，提出先采用SAC-IA粗配准算法求取变换矩阵，为ICP提供较好的初始位姿，再利用ICP算法进行精配准的方法。并在Ubuntu18.04.4 LTS 64位操作系统下对该方法进行验证，实验结果表明，本文提出的方法相较于传统ICP算法具有较好的匹配精度与收敛速度。

关键词：Kinect RGBD数据 点云分割 点云配准 ICP

Research on Work Positioning Method of Industrial Robot

Based on 3D Vision

Abstract

With the increasing demand for modern industry, machine vision has become a hot research topic. As an important part, point cloud registration provides a method for workpiece positioning of industrial robots. How to quickly and accurately achieve workpiece positioning has become the focus and difficulty in point cloud matching. This article is based on Kinect research on workpiece positioning method, the main work is as follows:

(1) Obtain RGBD data through Kinect to obtain workpiece point cloud. Pre-process the point cloud data: Use three filtering methods: pass-through, voxel, and statistics to greatly reduce the number of point clouds; A cluster segmentation algorithm based on Euclidean distance is used to segment the point cloud data into a single point cloud subset.

(2) Carry out the normal vector of the estimated workpiece point cloud, and use the Kd-tree spatial index method to accelerate the normal vector estimation for the problem of K neighborhood determination. Aiming at the problem of the large amount of calculation of the PFH feature descriptor, it is proposed to use the FPFH feature descriptor to describe the corresponding point pair to improve the calculation efficiency.

(3) Analyze the point cloud matching principle, aiming at the defects that the initial value of the ICP algorithm is high and easy to fall into the local optimum, it is proposed to use the SAC-IA rough registration algorithm to obtain the transformation matrix first, to provide a better initial pose for ICP Then use ICP algorithm for precise registration. The method is verified under Ubuntu18.04.4 LTS 64-bit operating system. The experimental results show that the proposed method has better matching accuracy and convergence speed than the traditional ICP algorithm.

Keywords：Kinect; RGBD data; point cloud segmentation; point cloud registration; ICP

目 录

摘 要 I

Abstract II

第一章 绪论 1

1.1 课题研究的背景及意义 1

1.1.1 课题研究的背景 1

1.1.2 课题研究的意义 2

1.2 国内外研究现状及趋势 2

1.2.1 国内外研究现状 2

1.2.2 课题发展趋势 3

1.3 本文研究的思路及主要内容 4

1.4 论文的章节安排 5

第二章 点云数据的获取 6

2.1 硬件平台 6

2.2 PCL点云开源库 6

2.2.1 PCL点云库简介 6

2.2.2 PCL库架构 7

2.3 图像获取与工件点云 7

2.3.1 RGBD图像 7

2.3.2 工件点云 8

2.4 本章小结 9

第三章 点云预处理 10

3.1 点云滤波方法 10

3.1.1 直通滤波 10

3.1.2 体素滤波 10

3.1.3 统计滤波 11

3.2 基于欧氏聚类的点云分割 11

3.3 本章小结 13

第四章 点云特征提取 14

4.1 点云法向量 14

4.1.1 点云法向量估计 14

4.1.2 点云空间索引 14

4.2 特征描述与特征提取 16

4.2.1 PFH特征描述子 16

4.2.3 FPFH特征描述子 17

4.3 本章小结 18

第五章 点云配准 19

5.1 SAC-IA粗配准 19

5.1.1 SAC-IA算法原理 19

5.1.2 SAC-IA实验结果 19

5.2 ICP精配准 20

5.2.1 ICP算法原理 20

5.2.2 ICP实验结果 21

5.3 本章小结 24

第六章 总结与展望 25

6.1 总结 25

6.2 展望 25

参考文献 27

致谢 28

第一章 绪论

1.1 课题研究的背景及意义

请支付后下载全文，论文总字数：22217字