本文为荷兰鲁汶大学（作者：Kris DEMARSIN）的博士论文，共150页。

CAD模型，即表示物体几何结构的计算机模型，在工业中是不可缺少的。CAD模型的创建通常基于逆向工程：3D扫描仪用于捕捉物理原型或成品零部件表面的点。从这样一组点（点云）构建CAD模型需要对点云进行分割，这样每个线段都可以由单个曲面片近似。因此，在本论文中，我们从点云提取出闭合的特征线，使得这些线将点云分割成不相交的点簇（段）。

我们区分了三种类型的特征线：非闭合对象点云边界处的边界边沿；将具有不同法向量的点区域分隔开的锐边；将具有不同曲率的点区域分开的平滑边缘。提出了一种基于点聚类和图形处理的特征线提取框架。

聚类的目的是将包含特征线的特征区域与其余的点云区分开来。因此，在框架中集成了每种类型要素线的特殊特性。应用图形处理步骤后，将特征区域简化为闭合的特征线。为了保证所有类型特征线的准确提取，我们提出了一种从整个点云中提取边界和锐利边缘的多层次方法，然后在由边界和锐边形成的子区域中检测平滑边缘。最后，我们讨论了用曲面片逼近点云的剩余步骤。提出了一种基于特征线提取的曲线网络生成算法，每个子区域及其边界曲线可以用一个曲面片来近似。

CAD models, i.e. computer models representing the geometry of an object,are indispensable in the industry. The creation of a CAD model is typicallybased on reverse engineering: a 3D scanner is used to capture points onthe surface of a physical prototype or a finished part. Constructing a CADmodel from such a set of points (point cloud) requires segmentation of thepoint cloud such that each segment can be approximated by a single surface patch. Therefore, in this thesis we extract closed feature lines from apoint cloud such that these lines segment the point cloud into disjoint pointclusters (segments).

We distinguish three types of feature lines: boundary edges at the borderof a point cloud of a non-closed object; sharp edges which separate pointareas with different normal vectors; and smooth edges that separate pointareas with different curvature. A framework is proposed to extract eachtype of feature lines, based on point clustering and graph processing steps.

The aim of the clustering is to distinguish the feature regions, which contain the feature lines, from the rest of the point cloud. Therefore, specialcharacteristics for each type of feature lines are integrated in the framework.After applying the graph processing steps, the feature regions are reducedto closed feature lines.To assure accurate extraction of all types of feature lines, we propose amulti-level method which extracts boundary and sharp edges from the entire point cloud. Smooth edges are then detected in the subregions that areformed by the boundary and sharp edges.At the end of this thesis we discuss the remaining steps to approximate apoint cloud with surface patches. An algorithm is presented to obtain acurve network, based on the extracted feature lines. Each subregion withits bounding curve can then be approximated by one surface patch.

1.  引言

2.  定义与几何特性

3.  研究背景与动机

4.  特征线提取的框架

5.  边界边沿提取

6.  锐边提取

7.  平滑边提取

8.  曲线网络构建

9.  结论

更多精彩文章请关注公众号：