场景点云的语义标注方法研究
发布时间：2018-10-14 09:12
【摘要】：近年来,随着机器视觉和人工智能技术的迅速发展,研究如何使机器具备空间场景理解和感知的能力,成为学者们研究的焦点。三维激光扫描测量技术的出现和发展为场景数据的获取提供了全新的技术手段,它采用非接触主动测量方式直接获取场景高精度三维点云数据,将机器对自然场景的理解和感知问题转换成对点云的数据处理问题。但由于三维激光扫描测量技术获取的点云数据具有海量、离散、缺乏语义信息等特性,因此,目前基于点云数据对场景的理解和感知仍是一项富有挑战性的工作。依赖复杂精细的生物视觉系统,人类天生具有对周围环境感知和理解的能力,但机器要实现对场景的感知和理解,则要艰难的多,需要依赖各种算法和模型。基于场景点云的语义标注方法,是当前提高机器对场景理解和感知能力的热点问题之一,也是本文的研究重点。该方法从整体上可以分为场景点云分割和场景点云标注两部分内容,其中点云分割主要是将整体点云分割成互不重叠的点云区域,形成各自独立的点云块单元,从而实现对场景中物体的 感知‖,但是要达到对场景点云理解的目的,就需要对分割出来的点云块进行分类标注。因此,本文的主要研究工作包括点云的分割和分类两部分内容,主要贡献如下:(1)为了实现地面点云的分割,本文提出结合关键点的随机采样一致性地面点云分割算法。此算法在原有算法的基础上增加多次迭代计算,统计获得包含最多模型点的平面模型,并以这些点为关键点进行平面拟合最终完成平面点云分割。改进后的算法比原算法对于有起伏的地面点云有更好的分割效果,提高了场景点云中地面点云的分割效果。(2)为了实现场景点云中建筑立面的点云分割,本文使用改进的区域增长算法来实现建筑立面点云分割。即在原有约束准则欧式最小距离准则的基础上,增加种子点和邻域点的法线角度这一判别条件,明显改变了建筑立面点云数据与地面点云数据交接部位的分割结果,使得这两部分的点云分割结果更加准确。(3)为了更好和更完整的实现场景中树木点云分割,本文在使用K-means聚类算法的基础上结合圆柱体拟合算法实现了整颗树木点云的分割。圆柱体拟合算法的加入,改变了只能分割树冠部分点云的局面,使得对于树木点云的分割更加完整和彻底。(4)借助分割获得的独立点云块获得高阶团构建场景点云的条件随机场模型,然后使用次梯度迭代算法和图割推断算法对条件随机场模型的参数进行学习和推断,再结合场景点云的特征向量,最终实现场景点云的语义标注。(5)在Visual Studio 2013开发平台下,使用C++编程语言,对相应的算法进行编程实现,最终达到对算法检验和验证的目的。点云语义标注方法的实现,具有重大的现实意义,大大加快了数字化城市建模的速度并为机器自动导航提供了判别依据,使机器获得了对空间场景感知和理解的能力,具有广阔的应用前景和应用价值。
[Abstract]:In recent years, with the rapid development of machine vision and artificial intelligence technology, the paper focuses on how to make the machine possess the ability of understanding and sensing space scene. the appearance and development of the three-dimensional laser scanning measurement technology provides a brand-new technical means for acquiring scene data, and adopts a non-contact active measurement mode to directly obtain the high-precision three-dimensional point cloud data of the scene, converting the understanding and perception of the machine to the natural scene into the data processing problem of the point cloud. However, it is still a challenging task to understand and perceive scene based on point cloud data because the cloud data acquired by three-dimensional laser scanning measurement technology has massive, discrete and lack of semantic information. Depending on the sophisticated biovision system, human beings are inherently capable of perception and understanding of the surrounding environment, but the machine needs to rely on algorithms and models to achieve the perception and understanding of the scene. Based on the semantic annotation method of the scene cloud, it is one of the hot issues to improve the understanding and perception ability of the machine at present, and it is also the focus of this paper. 璇ユ柟娉曚粠鏁翠綋涓婂彲浠ュ垎涓哄満鏅偣浜戝垎鍓插拰鍦烘櫙鐐逛簯鏍囨敞涓ら儴鍒嗗唴瀹，
本文编号：2270021
[Abstract]:In recent years, with the rapid development of machine vision and artificial intelligence technology, the paper focuses on how to make the machine possess the ability of understanding and sensing space scene. the appearance and development of the three-dimensional laser scanning measurement technology provides a brand-new technical means for acquiring scene data, and adopts a non-contact active measurement mode to directly obtain the high-precision three-dimensional point cloud data of the scene, converting the understanding and perception of the machine to the natural scene into the data processing problem of the point cloud. However, it is still a challenging task to understand and perceive scene based on point cloud data because the cloud data acquired by three-dimensional laser scanning measurement technology has massive, discrete and lack of semantic information. Depending on the sophisticated biovision system, human beings are inherently capable of perception and understanding of the surrounding environment, but the machine needs to rely on algorithms and models to achieve the perception and understanding of the scene. Based on the semantic annotation method of the scene cloud, it is one of the hot issues to improve the understanding and perception ability of the machine at present, and it is also the focus of this paper. 璇ユ柟娉曚粠鏁翠綋涓婂彲浠ュ垎涓哄満鏅偣浜戝垎鍓插拰鍦烘櫙鐐逛簯鏍囨敞涓ら儴鍒嗗唴瀹，
本文编号：2270021
本文链接：https://www.wllwen.com/shoufeilunwen/benkebiyelunwen/2270021.html
