【作者】 王丽辉；

【导师】 袁保宗；

【作者基本信息】 北京交通大学 ， 信号与信息处理， 2011， 博士

【摘要】 在计算机辅助几何设计、计算机动画、逆向工程、医学诊断、娱乐业等应用需求的推动下,三维点云数据处理技术受到越来越多的关注。人们可以通过很多种不同的方式获取现实世界中物体在计算机中的三维数据表示,本文的研究内容是将利用三维激光扫描仪获取的离散的三维点云数据,通过去噪光顺、特征检测、简化,获得能够更加准确表示现实世界中物体以及更适合应用的三维点云数据模型；并在此基础上对点云数据模型进行参数化以及曲面重建,获得三维物体的参数化、曲面和多边形模型描述,以便将其输出到虚拟环境中,完成从真实世界到虚拟世界的转换。本文取得的主要成果总结如下：1.提出了两种三维点云数据去噪算法。第一种是结合点云权重模糊C均值聚类算法(PWFCM)和点云双边滤波方法(PBF)的算法,它利用噪声的属性,将噪声分为大尺度噪声和小尺度噪声,将大尺度噪声直接去除,小尺度噪声移向PWFCM聚类中心并进行一次点云双边滤波光顺,避免了迭代计算,可以较好地保持模型的尖锐特征。第二种算法,适用于噪声环境复杂、数据点集规模较大的情况,它先利用点云边界检测去噪算法(PBD)检测并去除噪声,再对剩余的大尺度噪声,利用点云权重模糊C均值聚类算法(PWFCM)算法直接去除,可以减少数据量并可避免滤波产生的过光顺。2.提出了一种基于曲率和密度的三维散乱点云数据模型的特征点检测算法(CDFD)。该算法在特征检测时,利用三维点云数据模型的几何局部信息,包括邻居点的平均距离、邻居点法向夹角和曲率,来估计数据点的特征参数,然后利用数据点密度和模型到中心点的最大距离确定一个阈值,特征参数大于阈值的点被判定为特征点。该算法在不同的数据集上都能得到很好的实验结果,为模型简化和曲面重建提供比较好的基础。3.提出了一种模型均匀采样方法,该方法利用特征检测的结果,将三维点云数据模型的特征点和非特征点分离出来,对非特征点模型的球面参数化结果进行Isocube球面采样,可以得到比较均匀的采样结果。定义不同的采样率,就能得到非特征点模型多分辨率、均匀的简化模型,将模型特征点进行保护,不参加简化算法,可以最大程度的保留模型的尖锐特征。4.提出了一种三维点云数据模型的几何图像参数化方法。将点云模型进行球面参数化、正八面体参数化以及正八面体展开,得到三维点云数据模型的二维平面参数化结果,也就是几何图像模型。这样的参数化方法不需要进行多边形化和曲面拟合操作,因此计算量小,容易实现,而且参数化模型的几何变形很小。将三维点云模型转化到二维的几何图像模型后,很容易实现对几何图像进行重采样和Morphing操作,并逆映射回三维点云模型。5.以第三章点云模型的特征检测结果和第四章Isocube点云模型的简化结果为研究对象,提出了利用单尺度和多尺度的紧支撑径向基函数对点云模型插值,可以实现多分辨率的曲面插值的重建方法,该方法利用共轭梯度算法求解线性方程组可以减少计算量,提高计算速度。实验结果表明此方法对不同数据量的模型都可以得到很好的曲面重建结果。6.构建三维点云数据处理的实验系统,该系统实现了从真实世界物体形状信息中重建出相应的计算机内部描述的虚拟景物模型。系统采用平台方式,将三维点云数据处理中的关键技术及成果转化为若干层次的、相对独立的功能化模块,通过功能模块的有效、有机集成,实现三维点云数据处理系统的各项任务,从而为三维点云数据处理的进一步深入研究与完善奠定基础,又有助于三维点云数据处理研究的应用。更多还原

【Abstract】 In computer-aided geometric design, computer animation, reverse engineering, medical diagnosis, entertainment and other applications, the processing techniques of 3D point cloud data become more and more attentive. People can use different ways to get three-dimensional computer data of real-world objects. The content of this research is the use of 3D laser scanner to obtain the discrete 3D point data cloud and by smooth-ing, de-noising, feature detection, simplification to get more accurate and suitable data model for representation of real world objects. Then on this basis, the parameterization and surface reconstruction for 3D point cloud model are completed to obtain the para-meters, surfaces and polygonal model description of objects. Finally we output the re-sult to virtual environment and complete computation from real world objects into the virtual world reality.Main contributions of this thesis can be summarized as follows:1. Two 3D point cloud data de-noising algorithms are proposed. The first is point cloud weighed fuzzy c-means clustering (PWFCM) and point cloud bilateral filter me-thod(PBF) algorithm. We define the noise as large-scale noise and partly smooth small-scale noise. Large-scale noise is deleted directly and the small-scale noise will be moved to near the clustering center, then the remainder small-scale noise is smoothing by point cloud bilateral filter method (PBF). The second algorithm is for more complex cases. the data are de-noised first by point boundary detection (PBD) method, which detect the boundary point as noise and delete it, then delete the large-scale noise by point cloud weighed fuzzy c-means clustering (PWFCM). The algorithms proposed can decrease the amount of data and avoid over-smoothing.2. We propose a curvature and density based feature point detection method (CDFD). A new feature parameter is defined which considers the average distance and the normal angle between the point and its neighboring points and point curvature pa-rameter. This parameter shows local geometry information. We define also a feature threshold from data density and maximum distance of data points. Then the feature points could be recognized when its density parameter is bigger than the threshold. Ex-perimental results show that the new approach can detect the feature points accurately for different 3D scattered point data cloud models and it can provide the good model for further simplification and surface reconstruction.3. A uniformly sampling method for 3D cloud data is proposed. After the separa-tion of feature and non-feature data, we project the non-feature data of model to a sphere by using Isocube sphere and uniform sampling. With different sampling rate, we can get multi-resolution and uniform simplification model. All feature points are kept the same, so the sharp information of 3D point cloud may be retained well.4. We propose a geometry image parameterization algorithm for 3D point cloud data. Firstly, we do spherical parameterization, octahedron parameterization and unfold the octahedron to 2D plane parameterization and get a geometry image of 3D data. This operation does not need to do trianglization and surface fitting. So it is fast and easy to compete. From the operation, we transformed the 3D model to the 2D model. Then the re-sampling and morphing of the geometry image may be applied, and inverse mapping back to 3D point cloud model.5. On the basis of feature point detection and Isocube simplification of 3D cloud data, a multi-resolution surface reconstruction method is proposed, which uses the sin-gle scale and multi scale Compactly Supported Radial Basis Function interpolation. We use conjugate gradient algorithm for solving linear equations to reduce the compu-tational load and improve computing speed. The surface reconstruction result is good.6. An experiment system for 3D point cloud data processing is established. The system may reconstruct realistic parametric description of 3D objects for virtual envi-ronment usages. The platform can fulfill different 3D point cloud data processing tasks through effective, organic integration of hierarchical and relatively independent func-tional modules of data processing, which lays a solid foundation for the further in-depth study in 3D point cloud data processing research.更多还原