本文为德国慕尼黑工业大学（作者：Anas Al-Nuaimi）的博士论文，共156页。
由于众多技术的进步，3D传感变得越来越精确和廉价。目前已有的不同类型的三维传感技术，包括激光雷达（LiDAR）、飞行时间（Time-of-Flight）和多视点立体（Multi-View Stereo）。尽管3D传感器仍然相对昂贵，并且需要专业知识才能操作，但正是微软Kinect的发布使得精确和廉价的3D传感成为现实。它的成功伴随着计算机视觉研究社区的许多新发展，这使得许多新的应用成为可能。
三维传感器以三维点集合的形式捕获扫描，称为点云。如同照相机的图像一样，点云是视点相关的局部环境扫描。然而，与图像不同，可以将多次扫描组合成更全面的3D地图。将点云组合成更全面的点云，需要将它们沿着扫描的常见的特定形状进行对齐，这就需要揭示它们的局部坐标系之间的相对变换。研究界已经开发了不同的技术来处理点云配准问题。随着将多次扫描配准到更全面3D模型中的可能性，部分研究团体将重点放在开发和使用配准技术在3D中映射城市的室内和室外空间。这种3D地图对于许多新兴应用是必不可少的，包括机器人导航和环境交互、设施管理、文化遗产归档、虚拟和增强现实等等。
本文研究了点云配准问题，重点研究了三维测绘和室内定位的应用。本文主要有四个方面的贡献。贡献1涉及使用最流行的配准方法，即，使用特征匹配的基本配准理论。从应用的角度来看，贡献2的级别更高，因为它处理了运动LiDAR扫描的非刚性变形及其对扫描匹配的影响，扫描匹配是从单次扫描中组装成3D地图的过程中实现的。为此，利用了贡献1中设计的过滤器。在贡献3中，提出了一种专门针对由扫描匹配产生的3D地图的配准而定制的算法，使得可以从较小的3D地图中组装更全面的3D地图。在贡献4中，综合3D地图被用作参考点云，通过Kinect获得的对象扫描被配准到参考点云上，以便进行非常精确的位置检索。
Due to numerous technological advancements,3D sensing is becoming more accurate and affordable. Different types of 3Dsensing techniques exist including LiDAR, Time-of-Flight and Multi-View Stereo.While 3D sensors have been relatively expensive and require specialistknowledge to operate, it is the launch of the Microsoft Kinect that madeaccurate and cheap 3D sensing a reality. Its success was accompanied by manynew developments from the computer vision research community which enabled manynew applications. 3D sensors capture scans in the shape of collections of 3Dpoints, called point clouds. Just like images of cameras, point clouds are viewdependent partial scans of the environment. Unlike images, however, multiplescans can be assembled into more comprehensive 3D maps. Assembling point cloudsinto more comprehensive ones requires aligning them along distinctive shapesthat are common to the scans. This necessitates uncovering the relativetransformation between their local coordinate systems. The research communityhas developed different techniques to deal with the point cloud registrationproblem. With the possibility to register multiple scans into morecomprehensive 3D models, it is no surprise that part of the research communityhas focused on developing and using registration techniques to map urban indoorand outdoor spaces in 3D. Such 3D maps are essential to many emergingapplications including robot navigation and environment interaction, facilitymanagement, cultural heritage archiving, virtual and augmented reality and manymore. This thesis deals with the topic of point cloud registration with specialemphasis on the applications of 3D mapping and indoor localization. Fourcontributions are presented. Contribution 1 deals with the fundamentalregistration theory using the most popular registration approach, namely usingfeature matching. Contribution 2 is more high level from the applicationperspective as it deals with the non-rigid deformations of moving LiDAR scansand its impact on scan matching which is performed to assemble 3D maps from theindividual scans. For that the filter presented in contribution 1 is deployed.An algorithm specifically tailored for the registration of 3D maps generatedfrom scan matching such that more comprehensive 3D maps can be assembled fromsmaller ones is presented in contribution 3. In contribution 4, thecomprehensive 3D maps are used as reference point clouds against which objectscans obtained with the Kinect are registered for very accurate locationretrieval.
1 引言
2 项目背景
3 基于GA-LMS的点云配准
4 扫描偏斜及其对扫描匹配的影响分析
5 大型室内点云的解耦旋转平移对准
6 基于三维形状匹配的室内定位检索
7 结论与未来研究展望
附录 指数映射归一化常数
下载英文原文地址：
http://page2.dfpan.com/fs/cl4c9jf2b2d1b2f9160/
更多精彩文章请关注微信号：
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
