【作者】 黄海明；

【导师】 刘金刚；

【作者基本信息】 中国科学院研究生院（计算技术研究所） ， 计算机应用技术， 2005， 博士

【摘要】 本文从空间多点信息采集处理技术及其在虚拟现实中的应用出发,针对目前光学运动捕捉所存在的运动跟踪与重构的质量和速度等各种问题展开了研究,包括提高跟踪质量的策略和改善重构水平的策略,并对虚拟人景交互合成的策略进行了深入的研究。根据运动捕捉的流程,从各个环节提高跟踪质量和重构效果,包括姿态参数的获取、参数有效性判别、虚拟人的建模与驱动、以及在虚拟现实环境中的合理保持一致性的策略,并实现了相关的算法。本文首先综述了运动捕捉的意义,人体运动捕捉数据的不同应用,分析了对运动跟踪质量造成影响的各种因素并讨论了其中的一些需要进一步研究的问题。同时对人景合成的策略进行了介绍,还介绍了一些可以利用运动捕捉的虚拟人应用系统。本文针对与表演者匹配的marker-skeleton模型获取的相关问题展开研究,定义了适合运动的人体结构,研究并确定了适合的marker个数和粘贴位置。由于关节中心的获取是skeleton模型获取的关键问题,本文研究了各种关节中心获取算法,通过实验从算法精度和复杂度两个方面分析了各种算法的优劣,给出了marker权值、搜索优化、异点去除以及改善样本集等策略,最终实现了一种精确的关节中心定位算法。在此基础上,根据本系统的具体情况,进行局部关节子系统的调整,构造了Full-Marker-Joint关系模型来进行关节中心的快速定位,以此来生成所有帧的关节中心。接着在分析了Herda假设的基础上,提出了经验解析算法与全局能量最小搜索的混合方案,最后给出了完全marker-skeleton模型的生成结果。本文对人体运动姿态的运动参数获取方法展开了研究,分析了现有的多种姿态获取方法,包括运动学几何解析法、逆向运动学解析法、能量函数法和动力学方法。接着详细阐述了基于较少marker的无skeleton的运动参数解析法。为了解决分级解析几何关节参数求解所导致的最终重构的marker重合度不佳的情况,并避免局部收敛问题,最终本文提出了一种基于粗姿态和完全marker-skeleton模型的运动参数获取算法。本文分析了目前一般采用的运动参数有效性判别方法,包括刚体关系判别、幅度空间可达性判别、基于统计学的关节受限角判别、运动平滑度判别等等,然后提出了一种新的有效性判别策略——具有个性化关联性的运动参数有效性判别策略,通过生成、保存各关节的受限空间,将有效性问题转化为空间判定问题。最后对该策略进行了例证。在对信息补全与信息调整方法的研究中,我们特别考虑了速度变化的外插方法,并给出了有关的信息补全结果。同时针对Marker表面漂移问题,提出了一种基于常识针对表面漂移的关节参数补偿策略。在概述了目前虚拟人建模与驱动研究状况的基础上,本文简述了本系统虚拟人模型的建模方法,针对虚拟人匹配骨架的手工生成和配准所存在的随意误差,提出了一种基于OBB包围盒的虚拟人匹配骨架自动获取方法。然后文章阐述了常用的多肢体段重叠更多还原

【Abstract】 This dissertation begins with the development of spatial multipoint information collecting technology and its application in virtual reality, researches the problems on keeping coherence between motion tracking and reconstruction in current optical motion capture, including the strategy to improve the coherence and the reconstruction effect, and deeply researches the human-scene interaction synthesis. According to the flow of motion capture, the dissertation improves the tracking quality and reconstruction effect in various aspects, including obtaining posture parameters, judging the validity of motion parameters, modeling and driving virtual human, and the strategy of reasonably keeping the motion coherence in virtual reality environment, and then realizes correlative algorithms.Firstly, the dissertation summarizes the significance of motion capture, describes the various applications of human motion capture data, analyzes various factors which affect the motion tracking quality, discusses some problems which need further study. At the same time, the dissertation introduces the strategy of human-scene synthesis, and gives some virtual human application systems that can utilize motion capture.This dissertation studies on correlative problems about obtaining marker- skeleton model matching the performer, defines human structure fit for motion demand, studies and decides suited number of marker and the position for sticking. Obtaining the joint centre is the key problem of obtaining skeleton model, this dissertation studies on various algorithms for obtaining joint centre, analyzes the virtue and flaw in algorithm precision and complexity degree by experiment. Gives out the strategy of setting the weight of markers, optimizing search course, wiping off abnormal points and improving the sample set. Finally realizes a kind of algorithm for accurate positioning of joint centre. On the above basis, according the actual characteristic of our system, adjusts the local joint coordinate system, constructs Full-Marker-Joint relation model to rapidly locate the joint centre in order to obtain all joint centres in each frame. Then the dissertation analyzes the hypothetic of L.Herda, brings out a mixed strategy utilizing experiential analytic algorithm and global energy minimization, and presents the marker-skeleton model as a result in the end.The dissertation studies on the methods for obtaining posture parameters of human motion, analyzes various existing methods for obtaining posture parameters, including kinematics analytic method, inverse kinematics analytic method, energy function method and dynamic method. And then particularly describes the motion parameters analytic method based on less marker in which skeleton model is needless. In order to avoid the bad更多还原