【作者】 孙金虎；

【导师】 周来水；

【作者基本信息】 南京航空航天大学 ， 航空宇航制造工程， 2013， 博士

【摘要】 随着经济的发展和市场竞争的激烈，产品的生命周期不断缩短，用户对产品的需求日趋个性化。逆向工程是实现产品快速开发的途径之一，而通过逆向工程扫描测量获得的首先是点云模型，点云的分割与融合是基于点云模型进行产品快速与个性化设计的重要手段。“分割”是将产品局部细节数据从点云模型中分离提取出来，“融合”是将一个点云数据与其它点云数据拼接形成新的产品造型。本文研究点云模型分割与融合关键技术，主要的研究内容和成果如下：为了建立点云中点的局部拓扑关系，研究了点的邻域搜索技术，提出了一种基于立方体栅格的点云模型K邻域搜索算法。采用“直接搜索最外层栅格”对现有的逐层搜索算法进行改进，采用“动态缩小空间球半径”对现有的空间球算法进行改进，并引入“点数阈值”对这两种改进算法进行了组合，确保了栅格边长取不同值时K邻域搜索算法都具有较高的效率。点云的法矢对点云的分割和融合具有重要影响，为了得到全局一致（朝外）的法矢，提出了基于改进最小生成树的法矢调整及其奇异情况处理算法，确保了法矢调整结果的正确性，提高了法矢调整效率。在遍历方式和扩散方式上对传统最小生成树的生成过程进行了改进，提高了生成效率。奇异情况处理包含了对垂直法向和相邻曲面的处理，垂直法向的处理方法是增加邻域范围，相邻曲面的处理方法是剔除歧义邻域点。在剔除歧义邻域点时，给出邻域增长法、法向剔除法和组合法三种方法，其中组合法综合了邻域增长法和法向剔除法的优势，剔除效果最佳。为了获得点云模型中的特征分割线，提出了一种基于Snake的点云模型分割边界提取算法。创建点云Snake的气球力模型，将气球力由二维图像Snake推广到三维点云Snake中，用于设计点云Snake的外部能，确保初始轮廓远离目标轮廓时能够正确收敛。通过分步、分段、参数设定等方法提高了Snake的收敛效果。算法解决了点云Snake模型对初始位置敏感、受噪声影响以及受其它特征影响等问题，为点云Snake模型的进一步研究打下了基础。应用改进最小生成树实现了分割区域内部点的提取。在提取前，为了确保最小生成树能正确终止，算法将分割边界由线状扩展为带状。提取后，算法将带状边界拆分到各个区域中。算法能够有效避免过分割和欠分割，能够生成光顺分割边界，与已有分割算法Level Set相比具有较高的效率。研究了点云模型的融合技术，提出了一种点云模型融合过渡算法。采用径向基函数生成过渡曲面；利用基于Snake和改进最小生成树的区域分割算法提取过渡区域点云，操作简单且能进行精确分割；采用局部优化投影法对过渡区域进行均匀重采样；分析了影响过渡区域效果的因素并给出相应的处理方法。更多还原

【Abstract】 With the development of economy and the drastical competition in the market, lifecycle ofproduct is shorter and user demands are more personal. Reverse engineering is one of the methodsachieving rapid product development. Point cloud is original result got by reverse engineeringscanning. Point cloud segmentation and fusion is one important mehod for rapid and personal productdevelopment. Segmentation means extracting details from point cloud, and fusion refers to combiningone point cloud with others into new product model. Both technologies are studied. The main researchachievements are as follows:In order to establish the topology relations of local points, a K-nearest neighbors searchingtechnology based on cubes is researched. The layer-by-layer search method and the spatial spheremethod are improved by searching outermost layer directly and reducing the radius of spatial spheredynamically respectlvely. Then a threshold of point number in one cube is applied to synthesize thesetwo improved search methods and ensure the high efficiency for different cube side lengths.The normal vector of point cloud has great influence on point cloud segmentation and fusion. Inorder to get global consistent outward normal vector, a normal adjustment algorithm based onimproved minimum spanning tree and singular handling is proposed to adjust normal correctly andincrease efficiency. Minimum spanning tree is improved in traversal patterns and spread patterns toget higher efficiency. Two singular cases, which mean perpendicular normals and close-by surfaces,are considered. Increasing the neighbors region is used to handle perpendicular normals and removingambiguous neighbors is used to handle close-by surfaces. Three methods, which refer to neighborsincreasing method, normal points removing method and combination method, are given to removeambiguous neighbors. Among these, combination method takes advantages of other two methods andthus achieves best remove effect.For the purpose of getting feature segment line, the Snake model of point cloud is proposed andapplied to extract segmentation boundary of point cloud. The application of balloon force is extendedfrom2D image to3D point cloud in Snake to design external energy. Snake model can convergecorrectly even if the initial contour is far away from the target contour when using balloon force.Substep Snake, subsection Snake and parameter setting are involved to get better convergence effect.The proposed model is insensitive to initial position and can be less affected by noises and otherfeatures. It lays the foundation for further studies. Improved Minimum Spanning Tree realizes extraction of interior points. For proper finish ofimproved Minimum Spanning Tree, segment line is expanded towards both sides to bandedsegmentation boundary before extraction. After extraction, banded segmentation boundary is split tosingle regions. The algorithm can avoid over segmentation or under segmentation and generatesmooth segmentation boundaries. Compared with the Level Set segmentation algorithm, the algorithmof this paper can segment point cloud more efficiently.The transition surface creation algorithm is presented to research point cloud fusion technology.Radial basis function is adopted to construct transition surface. Transition region, the desired part oftransition surface, is generated by the point cloud segmentation algorithm based on Snake model andimpromved Minimum Spanning Tree. Not only is it easy to operate, but also the segmentation result isaccurate. The locally optimal projection algorithm is applied to uniform resample for transition region.Factors affecting transition region are analyzed to get corresponding dealing methods.更多还原