【作者】 张娜；

【导师】 孔德明；

【作者基本信息】 燕山大学 ， 检测技术与自动化装置， 2020， 硕士

【摘要】 逆向工程技术与计算机图形学技术不断发展,在复杂曲面结构模型的设计、制造等多个阶段得到了广泛应用,新技术的引入加速了模型的数字化进程。数字化处理后的模型可用于特征分析、形状检测,实现由实体模型到设计图纸的过程。其中,点云分割是逆向设计制造过程中的关键内容,对物体分类、目标识别、三维重建等任务至关重要。现阶段已有多种分割方法成功应用于复杂曲面结构模型,但由于模型种类复杂多样,包含的一些自由曲面边界难以界定,加之扫描设备获取到的点云数据中夹杂着噪点,对点云数据处理造成一定的干扰,因此实际分割实验中普遍存在部分特征区域识别效果欠佳的问题。为了提高分割的完整度,提出一种基于多视角区域生长的分割方法,主要研究内容包括:(1)针对复杂曲面结构模型数据量大、复杂度高的特性,提出一种基于法向量方向相异性原则的模型初分类方法。利用G2S(Gabriel2 Simplex)准则构建复杂曲面结构模型的三角网格结构,获取其表面的拓扑特征,解算并校正拓扑结构中每一三角面片的法向量信息,依据不同类间法向的指向差异将模型划分为不同类别的子集合。(2)针对单一视角下特征信息缺失与传统距离图像的缺陷,提出一种多视角距离图像的生成方法。在选取视角下将各类点云垂直映射到二维平面上,进行格网划分与网格赋值处理,实现三维点云到二维图像的转化,并引入基于图像形态学的平滑算法削弱图像中尖锐区域,获得不同视角下的距离图像。(3)针对传统区域生长算法自动化程度低、分割完整度不足的情况,提出一种改进的区域生长算法。借助多视角距离图像计算三维空间中不同特征区域内的种子点,采用网格法向量偏移角度和距离约束的区域生长算法识别不同特征面,并利用KNN(K-Nearest Neighbor,KNN)算法剔除离群点,在多种复杂曲面结构模型上均可获得不低于80%的合理分割结果。更多还原

【Abstract】 With the development of reverse engineering technology and computer graphics technology,it has been widely used in the design and manufacture of model with complex curved surface structure.The introduction of new technologies has accelerated the digitalization of the model.The digitized model can be used for feature analysis and shape detection to realize the process from solid model to design drawing.Among them,point cloud segmentation is the key content in the process of reverse design and manufacture,which is very important for the tasks of object classification,target recognition,3D reconstruction and so on.Many segmentation methods have been successfully applied to model with complex curved surface structure.However,because of the complexity and variety of model,it is difficult to define the boundary of some free-form surfaces.In addition,noise is mixed in the point cloud data obtained by scanning equipment,which causes some interference to the point cloud data processing.Therefore,the problem of poor recognition effect of some feature regions is common in the actual segmentation experiment.In order to improve the completeness of segmentation,a segmentation method based on multi-view region growing is proposed.The main research contents include:(1)Aiming at the characteristics of large volume of data and high complexity of the model,an initial classification method based on the principle of normal vector differences is proposed.The triangular mesh structure of model with complex curved surface structure is constructed based on G2S(Gabriel2 Simplex)criterion to obtain the topological features.The normal vector information of each triangular in the topology is calculated and corrected,and the model is divided into subregion of different categories based on the difference of normal vector direction between different categories.(2)Aiming at the lack of feature information from a single view and the defects of traditional distance image,a multi-view distance image generation method is proposed.All categories of point cloud are mapped vertically to 2D plane from the selected angle of view for grid division and grid assignment.The smoothing algorithm based on image morphology is introduced to weaken the sharp areas in the image,and then distance images under different angles of view are obtained.(3)In view of the low automation and insufficient segmentation integrity of traditional region growing algorithm,an improved region growing algorithm is proposed.Seed points in different feature regions in 3D space are calculated based on multi-view distance image.The region growing algorithm with constraints of grid normal vector offset angle and distance is used to identify different feature surfaces,and the KNN(K-Nearest Neighbor,KNN)algorithm is used to eliminate off-group points.The reasonable segmentation result of not less than 80% can be obtained on various models with complex curved surface structure.更多还原