基于形状分析和条件随机场的三维点云分类

CASIA OpenIR > 毕业生 > 硕士学位论文

	基于形状分析和条件随机场的三维点云分类
	刘欣莹
	2017-05
学位类型	工学硕士
中文摘要	随着三维扫描技术的不断深入发展，点云数据的获取变得简单快捷。目前，通过各类机载、车载、手持扫描仪等设备均可高效地获取高精度大范围的点云数据，这为点云的分析和重建创造了良好的条件。在点云处理的一系列过程中，点云分类是不可缺失的重要组成部分，也是当前计算机科学相关领域的研究重点和热点。近年来，有关点云分类的研究取得了较大进展。但是，对于诸如道路桥梁、树木等包含相互交错遮挡的复杂三维点云数据，现有方法的性能仍难以达到相关应用需求。为此，本文将针对地形地貌和多种树木等点云数据进行形态特征统计分析，并在此基础上开展点云分类研究工作。本文的主要研究工作和贡献如下： 1. 提出局部形状特征概率混合的半自动三维点云分类方法。具体地，首先利用形状概率统计特征对点云数据进行计算，再利用解析曲面模拟采样点集，对近邻距离、近邻四面体体积、近邻法向量差异度和主曲率方向差异度这四个形状特征进行统计分析和比较；然后，针对上述四种特征，提出了一种基于概率混合的多特征融合方法，从而弥补单一特征在分类判别能力上的不足；在此基础上，提出基于主曲率阈值投票的快速点云聚类方法。本文实验结果验证了所提方法的有效性。 2. 提出了一种基于条件随机场的点云分类方法。具体地，基于点的最近邻四面体体积、高斯曲率、点对势能等局部形状特征，首先利用点云的位置坐标信息进行条件随机场模型构建；然后通过正则化对数条件概率（Regularized Log Likelihood）构造似然函数，在总体分类精度最优准则下利用蒙特卡洛方法进行参数推断；在此基础上，进一步提出一种基于最大化类间加权准确率的参数学习准则，该准则强调各个类别的准确率，可使得最后分类结果中每一类的准确分类达到一个平衡。在一系列树林点云数据的实验结果表明，所提方法能够较大地提高树木点云的分类正确率。
英文摘要	With the continuous development of three-dimensional scanning technology, point cloud data acquisition is becoming easier and faster. Employing various types of airborne onboard and handhold scanners, high-precision large-scale point cloud data can be obtained efficiently, which offers excellent esources for point cloud analysis and reconstruction. For point cloud data processing, point cloud classification plays a vital role, and is also a hot topic in the field of computer science. Although the studies on cloud point classification had made great progress in recent years, traditional classification methods cannot meet the requirements for accurate analysis and classification of the complex 3D point cloud data that intersects each other such as trees, roads or bridges. In order to solve this problem, we focus on how to extract the shape feature from terrain/topography data and a variety of trees point cloud and make corresponding statistics, and classify these data into different classes for further applications. The main contribution of our work are described as follows: 1. A semi-automatic three-dimensional point cloud classification method based on mixing the local shape feature probabilities. Specifically, we first make statistical analysis of the local shape feature probabilities are made on the point cloud data, and then use analytic surfaces to approximate the sampling point set. Next, four local shape features including the nearest neighbor distance, the neighboring tetrahedral volume, the nearest neighbor normal vector difference and the principal curvature direction difference degree are extracted and compared with each other to evaluate their ttributes. In view of the above four features, a multi feature fusion method based on probabilistic mixing is proposed to compensate for the shortage of using single feature in classification and discrimination. Correspondingly, the principal curvature threshold voting method is developed to cluster the point cloud data with fast computation speed. Experimental results validate our proposed method. 2. A point cloud classification method based on the conditional random field. Specifically, based on the local shape features of each point including the nearest neighboring tetrahedron volume, the Gaussian curvature and the point potential, a conditional random field (CRF) is constructed by using the spatial coordinate information of the point cloud. Then, Regularized Log Likelihood is utilized to construct the likelihood function and the Monte Carlo method is applied to optimize the parameters under the criterion of the optimal overall accuracy. Based on our constructed CRF model, a new parameter learning criterion is further proposed to maximize the weighted interclass accuracy. Our criterion pays more attention to the accuracy of each category, so that we can trade off the correct classification of each category in the final classification results. A series of experiments conducted on the forest point cloud data sets validate that our classification method can greatly improve the classification accuracy of the tree point cloud.
关键词	点云处理局部形状特征点云分类特征提取条件随机场
学科领域	计算机应用技术
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/14753
专题	毕业生_硕士学位论文
作者单位	中科院自动化研究所
第一作者单位	中国科学院自动化研究所
推荐引用方式 GB/T 7714	刘欣莹. 基于形状分析和条件随机场的三维点云分类[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
Thesis.pdf（45094KB）	学位论文		限制开放	CC BY-NC-SA