三维散乱点云的特征提取方法研究

发布时间：2018-11-06 19:04

【摘要】：随着三维测量技术的不断发展,通过数字扫描设备能够有效的获取现实世界中物理物体的高精度表面模型,并成为获取3D点云数据的主要手段。三维点云模型已经被广泛应用于模式识别、三维重建、模型分割等领域,而特征提取作为三维点云模型处理的底层技术成为图像领域的研究热点。本文在总结特征提取技术国内外研究现状的基础上,将马尔科夫随机场(Markov Random Field,MRF)模型应用在该领域,分别从建立典型的MRF模型和提取显著特征点建立MRF模型两方面给出研究思路和求解框架。本文研究内容主要包括:1.提出了基于马尔科夫随机场的散乱点云全局特征提取算法。该算法基于经典的MRF模型,通过观察点云分布的直方图拟合高斯分布建立模型,根据贝叶斯估计将先验问题转化为最大后验概率的求解,进一步推导为随机场最小能量的求解,归约简化得到目标函数,求解该函数并提取特征点。该算法针对传统算法存在大量的人工调参及阈值设置的问题,灵活的融合了典型MRF模型的自适应性,有效避免了传统算法的弊端,并提高了算法的自适应性和时间效率。2.提出了基于显著特征点的散乱点云特征提取算法。该算法的核心思想是对典型的MRF模型进行改进,和前面算法的主要区别是对随机场模型的建立方法:通过构造点云的显著度函数计算散乱点的显著度,结合点间测地距和显著度构建Reeb图,提取显著特征点,根据点到显著特征点以及中心点的距离求MRF随机场的联合密度函数。该算法继承了自适应性优点,同时也避免了上面算法存在初始阈值的设置及点云数据高斯拟合的问题,使点云特征提取完全跳脱出传统曲面曲线拟合和特征参数设置的思维禁锢。3.这两个特征提取算法主要应用于秦兵马俑碎片的虚拟复原项目中。实验结果表明本文算法均能够有效提取出兵俑的特征,相对于传统算法具有自适应性和高效性,为后续兵马俑虚拟复原工作奠定了基础。
[Abstract]:With the development of 3D measurement technology, the high precision surface model of physical objects in the real world can be obtained effectively by digital scanning equipment, and it becomes the main means to obtain 3D point cloud data. 3D point cloud model has been widely used in pattern recognition, 3D reconstruction, model segmentation and other fields. Feature extraction as the bottom technology of 3D point cloud model processing has become the focus of image research. On the basis of summarizing the research status of feature extraction technology at home and abroad, this paper applies Markov Random Field (Markov Random Field,MRF) model to this field. The research ideas and solution framework are given from two aspects: establishing typical MRF model and extracting prominent feature points to establish MRF model. The main contents of this paper are as follows: 1. A global feature extraction algorithm for scattered point clouds based on Markov random field is proposed. Based on the classical MRF model, the model is established by fitting Gao Si distribution with histogram of observation point cloud distribution. According to Bayesian estimation, the priori problem is transformed into the solution of the maximum posterior probability, and the solution of the minimum energy of the random field is further deduced. The objective function is obtained by reduction, the function is solved and the feature points are extracted. Aiming at the problems of manual parameter adjustment and threshold setting in traditional algorithms, the algorithm combines the self-adaptability of typical MRF model flexibly, effectively avoids the disadvantages of traditional algorithms, and improves the self-adaptability and time efficiency of the algorithm. 2. An algorithm for feature extraction of scattered point clouds based on salient feature points is proposed. The core idea of this algorithm is to improve the typical MRF model. The main difference from the previous algorithm is the method of establishing random field model: the salience of scattered points is calculated by constructing the salience function of point cloud. The Reeb map was constructed by geodesic distance and saliency between the combined points, and the joint density function of MRF random field was obtained according to the distance from the point to the significant feature point and the distance from the center point to the significant feature point. The algorithm inherits the advantages of self-adaptability and avoids the problem of initial threshold setting and point cloud data Gao Si fitting. The point cloud feature extraction completely jumps out of the traditional curve fitting and feature parameter setting. 3. 3. These two feature extraction algorithms are mainly used in the virtual restoration project of Qin Terracotta Warriors and horses fragments. The experimental results show that the proposed algorithm can effectively extract the features of the terracotta warriors, and is more adaptive and efficient than the traditional algorithms, which lays a foundation for the subsequent virtual restoration of the terracotta warriors.
【学位授予单位】：西北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】