点云特征线提取是点云模型重构的基础, 国内外对此从边缘检测、特征线跟踪和面域分析等方面展开了研究, 但由于存在模型多样性、点云数据噪声和不完整性、特征复杂性等问题, 看似简单的特征线自动化提取很难实现。从曲率突变点隐含了点云特征线这一论断出发, 借鉴图像处理中的区域分割和边缘检测思想, 提出了特征线提取中的聚类、细化、分段和排序方案。在具体实现中分别提出了基于连通区域聚类的备选点集分离算法, 基于局部影响区域腐蚀的点集细化算法, 以及基于组合搜索准则和主成分分析(PCA)双向搜索的特征线分支截断和排序算法。在对比实验中, 确定了算法关键参数曲率突变点比例w和方向夹角阈值θT的推荐值, 并与类似算法对比能提取更多的特征点; 在模型实验中, 简单几何模型的特征线提取正确率达到了100%, 复杂机械零件模型和艺术品模型的特征线提取正确率均达到了85%以上, 取得了预想的棱线和特征轮廓线提取效果。算法具有通用性和可扩展性, 通过程序优化可获得更好的特征提取效果。

Point cloud feature-line extraction is the basis of point cloud model reconstruction. Globally, the research on feature-line extraction from point clouds has been conducted in terms of edge detection, feature line tracking, and surface analysis. However, this seemingly simple operation is actually difficult to realize because of problems such as varied models, point cloud noise and imperfection, and feature complexity. Starting from the assertion that curvature mutated points imply the feature line of a single point cloud, this study presented a clustering, refinement, segmentation, and sorting scheme for feature-line extraction. This scheme was based on the idea of regional segmentation and edge detection in image processing. In the specific implementation of this scheme, this study proposed an alternative point set segmentation algorithm based on connection region clustering, a point set thinning algorithm based on locally influenced area corrosion, and a feature line branching truncation and sorting algorithm based on combinatorial search criteria and principal component analysis bidirectional search. Recommended values of the two key parameters, that was, curvature-mutated point ratio w and directional angle threshold θT, were determined by conducting a comparative experiment. The proposed method could also extract a greater number of feature points as compared to the similar algorithm. The anticipated contour extraction effect was achieved through model experiments, in which the accuracy ratio of the feature-line extraction for models of simple geometry and complex mechanical parts, as well as artwork models, are 100% and 85%, respectively. The algorithm has the characteristics of generality and extendibility, thus enabling an improved feature extraction effect to be obtained through program improvement.