基于基本形状体及其拓扑结构的点云场景物体识别方法研究

The point cloud has a strong expression ability of the complex scenes and object shape. However, the recognition of the scattered data points seriously hinders the application in robot navigation and machine vision. In this research, we propose a systematic method of the object recognition in point cloud scene. Firstly, taking the space topology for guidance, we construct the mechanism of object combination using Gauss mapping and spatial data structure and then an effective object combination library is established which provides support for the iterative formation and matching of object in the point cloud scene. Secondly, the topologic relations between the basic shapes are extracted to form an effective methodology for scene decomposition. Then, the iterative combination strategies of object are explored based on the object combination library and the combination and verification methods of object are also formed so as to realize the objects recognition in the scene. We also discuss and solve the problem of over-segmentation and under-segmentation in this work. Finally, a prototype software system is formed. The core purpose of the research is to develop the basic theories and techniques of 3D object recognition, realize the integration of digital geometry and pattern recognition, and further promote the wide application of the scattered data points in many fields.

点云对复杂场景和物体的"外形"具有强大的表达能力，但点云识别问题严重妨碍了它在机器人导航、机器视觉等领域的推广及应用。该课题以识别为目标，探讨和研究点云场景中物体对象识别的方法与体系。为此，我们以几何学中的基本几何形状体为依据，以空间拓扑学中的拓扑结构为导引，以高斯映射和空间数据结构为手段，拟建立基于基本形状体的物体对象组合构建机制，形成有效的组合对象库，为点云场景物体对象的迭代形成和匹配提供支撑；获得点云场景中基本形状体及其之间的拓扑结构关系，形成场景分解的有效方法体系；探索以组合对象库为参照依据的场景物体对象迭代组合策略，形成场景物体对象组合与验证方法，进而实现场景中物体对象的有效识别，同时探索解决分割中常见的过分割和欠分割这一关键问题。最后形成原型软件系统。该课题的核心目的是发展三维识别的基础理论与技术,实现数字几何与模式识别的融合,推动离散点集的广泛应用。

本项目结合实际应用，以三维场景点云数据为研究对象，探讨场景中物体对象的识别与重建。主要完成了：(1)提出了一种基于基本形状的点云场景分割方法。(2)提出了一种基于高斯映射的点云分割方法。(3)提出了一种基于结构的点云场景提取方法。(4)提出了一种基于结构的点云场景对象识别方法。(5)提出了一种场景中人脸姿态无关的识别方法。(6)提出了一种基于形状特征分类的点云场景重建方法。(7)提出了一种基于切片的复杂建筑物基本构件提取与重建方法。(8)提出了一种基于点云的曲面细分和亮度增强方法。该课题的研究为可视化、物体识别与理解和场景重建提供理论或方法上的依据。全面完成了计划任务书中的各项要求。