【作者】 孙晓东；

【导师】 张鸿宾；

【作者基本信息】 北京工业大学 ， 计算机应用技术， 2012， 博士

【摘要】 三维扫描技术可以快捷便利的获取三维模型，因此在逆向工程设计、仿真、医学训练、外科手术、人体测量及服装工程等领域都有广泛的应用。随着三维扫描技术的发展掀起了对三维扫描模型点云的研究热潮。点云的模型表示方法具有数据结构简单、存储空间紧凑和强大复杂表面局部细节表达的优势，因此直接以点云为对象的研究引起众多研究者的兴趣。然而由于人体测量学以及服装工业自身性质等因素以及全三维人体数据采集上的难度问题等导致了针对人体测量学和服装工业的专用点云处理算法发展却严重滞后无法满足人体测量学、三维服装CAD的需求。目前三维人体点云的处理研究上多依赖于通用点云模型算法或转化为网格模型进行处理。通用的点云算法具有普适性的优点，然而却很难满足人体测量学以及服装工业的专业需求。因此对人体点云展开针对性的研究非常必要。本文将集中研究人体点云，主要以全三维人体扫描仪获取的3D静态人体点云数据为主要研究对象，同时也对快速动态点云人体的获得方法做了研究。主要完成了对3D静态点云数据的分析：包括针对人体测量学和服装工业的需求对三维人体扫描云的修补、简化以及点云人体骨架提取，以及对近红外连续深度图序列计算得到的动态点云人体数据做初步研究，取得了相应的成果。本文的研究成果和创新点主要包括以下几个方面：1、提出一种具有全局光滑性的人体测量学专用的漏洞修补方法。算法快速有效，修补的点更符合人体测量学和服装工业研究的需求。修补的点更符合真实人体曲面特性。2、提出了保持人体特征的点云精简算法。利用点云二阶微分量识别特征点，并通过设定自适应权值来达到不同层次的精简。精简中可保持人测量学的关键特征点，并且在深度精简中不会出局部细小部位的衰退而造成的拓扑丢失现象。3、提出了一种在边界噪音高的原始扫描人体点云上提取骨架线的方法。使用分片Laplace谱嵌入与三维矩相结合的骨架线提取算法，降低谱嵌入的计算代价。算法计算稳定、对噪音具有良好适应性，并且在骨架线的中心性和算法的强壮性之间取得了较好的平衡。算法对人体的体型分类分析具有重要意义。还利用深度摄像头得到一种低成本、快速获取动态人体点云的方法。并使用动态点云信息作为先验指导，提出了一种对光线和纹理变化不敏感的人体头部追踪方法，算法具有一定的错误检测与自我错误恢复能力，可以较好克服追踪中的纹理和光线问题。本文的研究成果有利于促进服装三维CAD的发展，对人体测量学的体型分析与中国人体号型普查、制作符合中国人体的标准人台、个性化的服装定制等以及基于精确人体的动态服装展示等都有重要应用价值。更多还原

【Abstract】 3D scanning technology has found it application in inverse engineering, simulation, medical training, surgery, anthropometrics and garment engineering.3D scanned data have gain concentrate research with the development of3D scanning technology.Point cloud in contrasts to mesh based model have the advantage of no redundancy data gains the concentrate research. The point based3d human model is more feasible to obtain any level of anthropometrics information than the meshed one.There relatively fewer researches direct on point based3d scanned human body for anthropometrics which mainly relay on the general point cloud methods or turn the point cloud model to mesh for further analysis.Although the research on point based model had improved with the point representation geometry, there relatively few technology on point cloud of human body (HB) perfectly fulfills the needs of anthropometrics and garment industry. The research on specialized point cloud human body is lag behind. For example the customization of garment and the analysis (pressure etc.) of functional clothing on human body. virtual try on true3D human model and decimally accurate show of in3d garment designed pattern in sense of true accurate in stead of visual fidelity. The current research on point cloud is on general point cloud of garment engineer and the anthropometrics. How to adopts and improves the level-of-art methods on mesh based model or general point cloud model to scanned HB in order to meet the needs of anthropometrics or garment engineering.This paper is concentrated on the researches of point cloud of human body which aim to form effective specialized methods for anthropometrics or garment engineering.The main contributions of this work are on the followings:1. Presents a global smooth a filling holes strategy for HBS data. And the filled data are more faith to the real human body and could meet the need for anthropometrics.2. Proposes a simple and effective feature preserved simplification strategy for points cloud. It identifies feature points and then simplifies by self-adopted weight. During any level of simplification the important feature point of anthropometrics could preserved. With constraint to HBS simplification, the algorithm could preserve the anthropometrics key points effectively and efficiently even in sharp simplification rate. The simplification process is controllable and more feasible and used the feature set could extraction the longitudinal curves (in point set) of human for garment design at any level of detail with convenient, which could offer reliable data further shape analysis. The algorithms avoided the thin part attenuation which caused topological missing even in sharp simplification.3. Presents a skeleton curve extracted algorithm directly on scanned Human body point cloud with large boundary noises and gaps. The algorithm used the sliced embedding strategy to decrease computational costs, and it prevents the skeleton contraction off the central of model and the algorithm is insensitive to boundary noises. It is very important for analysis and classifies human body type.Also describes a low costs method to for generating dynamical human bodypoint cloud with depth serials. And finally presents human head tracking methodswith the dynamical point cloud as Priori knowledge guided. It has the ability oftracking error self-detected and recovery. It is insensitive to the texture of face eithercaused by illumination and different people.The achievements have benefits to3D computer aided garment design, anthropometrics for analysis of human body, which are very important for analysis Chinese body shape and form standard digital mannequin for Chinese body shape and also has benefit to dynamically show on accurate3D human body.更多还原