论文总字数：29123字

摘 要

本课题的研究内容是基于点特征直方图的点云拼接。物体的点云数据的获得由于技术限制往往不是全方位的，通常我们采用从单一视角拍摄某个物体，得到这个视角的模型数据，重复以上的操作，多视角进行拍摄，最后通过点云拼接技术得到研究对象的完整三维模型。由于所拍摄得到的点云数据量很大，如果直接原封不动采用的话会花费大量的时间计算，同时由于存在噪声离群点的干扰，反而也不一定能得不到好的结果。因此为了保证配准精读，同时减小计算量，先要对点云进行关键点提取。本课题使用的是ISS（Intrinsic Shape Signatures）关键点提取算法，它利用协方差矩阵建立模型，为了得到形容该点特征程度的特征值，采用奇异值分解的办法。最后通过设定阈值，满足判断条件的即为关键点。

在得到关键点后，计算关键点处的FPFH（Point Cloud Feature Histogram），得到所有关键点的统计特征，利用FPFH进行特征匹配，得到许多的关联点。同样的，实验表明特征匹配中有许多的错误匹配，若不去除会对配准精读有很大的影响。本文选用RANSAC（RANdom Sample Consensus）算法去除错误匹配后，计算出旋转矩阵，作用于源点云上完成初配准，使两片点云距离变近，模型上基本重合。以上为粗配准步骤，使得不同视角的点云变换到同一坐标系下，利用有重叠部分的两片点云进行一系列计算完成配准。之后利用ICP（Iterative Closet Points）算法实现精细配准，它假设点云中某点到另一点云中欧式距离最近的点作为匹配点，计算出点云的变换矩阵进行对源点云的坐标变换，作为下一次迭代的输入，一直这样迭代下去，直至收敛为止。对于ISS算法和RANSAC算法中要设置的阈值的问题，本文均采用一种自适应的方法进行设置，计算输入点云的分辨率，以其为一个单位大小附以相应的权重进行设置，能达到极好的效果。

关键词：点云配准 ISS关键点 FPFH特征 RANSAC算法 ICP算法

Point Cloud Registration Based On Point Cloud Feature Histogram

Abstract

The research content of subject is point cloud registration based on point feature histogram. The acquisition of point cloud data is often not omni-directional. Usually, we use a single view to capture an object, get the model data from this angle of view, repeat the above operation, and take multiple perspectives. Finally, we get the complete 3D model of the object through the point cloud registration technology. Because of the large amount of data taken from the point cloud, it will take a lot of time to calculate if it is used directly. At the same time, it can not get good results because of the interference of noise outliers. Therefore, in order to ensure the registration intensive reading and reduce the amount of computation, we need to extract the key points of the point cloud. The research uses the ISS (Intrinsic Shape Signatures) key point extraction algorithm, which uses the covariance matrix to establish the model. In order to get the eigenvalues of the point feature, the singular value decomposition method is adopted. Finally, by setting the threshold, it is the key point to satisfy the judgment condition.

After getting the key point, the FPFH (Point Cloud Feature Histogram) at the key point will be calculated, and the statistical characteristics of all the key points are obtained. A lot of correlation points are obtained by using the FPFH to match the features. Similarly, experiments show that there are many mismatches in feature matching. If not removed, it will have a great impact on registration and intensive reading. In this paper, RANSAC (Random Sample Consensus) algorithm is selected to eliminate error matching, and the rotation matrix is calculated, and the initial registration is completed on the source point cloud. The distance of two point clouds is close, and the model is basically coincided. The above is the rough registration step, which makes the point cloud from different perspectives to the same coordinate system, and uses a series of two point clouds with overlapping parts to perform a series of calculations to complete the registration. Then the ICP (Iterative Closet Points) algorithm is used to achieve fine registration. It assumes that the point in the cloud is the nearest point in the other point of the cloud as the matching point, and the transformation matrix of the point cloud is calculated for the coordinate transformation of the source point cloud. As the input of the next iteration, it is always iterated down to the convergence stop. For ISS and RANSAC algorithms, the research uses an adaptive method to set up the resolution of the input point cloud, which can be set with the corresponding weight for a unit size, and can achieve excellent results.

KEY WORDS: Registration, FPFH feature, ICP algorithm,

目 录

摘 要 I

Abstract II

第一章 绪论 1

1.1 引言 1

1.2 问题描述 1

1.3 点云拼接的研究现状 1

1.4 近年的迭代最近点算法改进策略 2

1.5 本文主要研究内容和结构 2

第二章 点云滤波算法的研究 4

2.1 引言 4

2.2 几种滤波算法 4

2.2.1 均值滤波算法 4

2.2.2 中值滤波算法 4

2.2.3 高斯滤波 5

2.2.4 双边滤波 5

2.3 离群点的移除 5

2.4 实验分析 6

第三章 点云关键点提取方法的研究 8

3.1 引言 8

3.2 关键点提取算法 8

3.2.1 NARF关键点 8

3.2.2 ISS关键点 8

3.3 关键点提取效果 9

第四章 基于点特征直方图的特征描述 13

4.1 引言 13

4.2 法线计算 13

4.2.1 法线理论 13

4.2.2 法线计算效果 14

4.3 FPFH描述子 15

4.3.1 PFH描述子 15

4.3.2 FPFH描述子 16

4.3.3 PFH和FPFH的区别 17

4.4 FPFH实验效果 17

第五章 特征匹配与误匹配去除 21

5.1 引言 21

5.2 特征匹配 21

5.2.1 搜索树 21

5.2.2 特征匹配 22

5.3 采样一致性算法 22

5.3.1 RANSAC随机采样一致性算法 22

5.3.2 LMedS最小中值方差估计算法 24

5.4 实验效果 24

第六章 旋转矩阵 28

6.1 引言 28

6.2 奇异值分解 28

6.3 四元数法 28

6.4 最小二乘法 29

6.5 实验效果 30

剩余内容已隐藏，请支付后下载全文，论文总字数：29123字