【作者】 瞿师；

【导师】 吴玲达；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2011， 博士

【摘要】 在虚拟环境中生成人体运动是计算机图形学和计算机动画的研究热点之一，能够广泛应用于游戏、影视、广告、军事等领域。由于人体结构的复杂性，生成自然逼真的人体运动是一项烦琐的工作，即使借助动画制作软件，生成一段运动也需要耗费大量的时间和精力。随着运动捕获技术的发展和运动捕获设备的普及，出现了一种新的运动生成方法，即基于运动捕获数据驱动虚拟角色的方法，该方法具有易于实现、生成的运动逼真度高等优点。但是，由于人体运动的多样性，有限的运动捕获数据无法满足广泛的实际应用的需求，基于现有运动生成新运动、提高现有运动的可重用性便凸显其重要性。本文研究的目的在于基于现有的人体运动数据，通过用户简单的交互式控制，生成自然逼真的新运动，以提高现有人体运动的可重用性，降低人体运动生成的复杂度。其中涉及的关键技术主要包括运动分割、运动的非线性降维、运动生成与编辑、运动连接、运动重定向等。本文综合运用统计学习、模式识别、数据挖掘、计算机图形学、最优化理论等学科理论对相关问题展开研究，对以上关键技术提出可行的解决方法。在理论研究和算法实践过程中，根据所研究的内容以及所要解决的问题，有针对性地提出了一些新思路和新算法，本文主要的创新点和贡献有：提出一种基于运动特征变化探测的运动自动分割方法。运动捕获设备获取的运动素材往往包含多个语义特征，在运动创作时一般是对其中的某个语义特征的运动片段进行素材重用，因此需要将包含多个语义特征的长运动分割为独立语义特征的运动片段。运动的语义特征属于主观认知范畴，为了避开对运动语义特征直接分析和建模的困难，将运动映射到低维特征空间，在特征空间提取运动几何特征并使之与运动的高层语义特征相对应，运动几何特征的变化能够反映运动语义特征的改变。该方法借助分析运动的几何特征来分析运动的高层语义特征，实现运动在语义层次的自动分割。提出快速自适应比例高斯过程隐变量模型，并提出一种基于该模型的人体运动数据降维及运动生成方法。针对人体运动数据维度高、不易直接分析的特点，通过对运动数据进行统计学习，实现非线性降维的同时获得了该运动姿势空间的概率描述，姿势概率作为姿势自然逼真程度的一个量化度量，概率越大表示姿势与样本运动越相似，因此越自然逼真。在给定末端约束条件下求取满足约束的、同时概率最大的姿势，作为逆向运动学的解，克服了传统逆向运动学算法计算烦琐、效果不逼真的缺点。在姿势生成的基础上，给出了两种运动生成方法，即基于运动轨迹编辑的运动生成和基于关键帧编辑的运动生成。相对于比例高斯过程隐变量模型，该模型具有更快的收敛速度和更高的收敛精度，同时能够自适应运动编辑的方向，有效地扩大了运动的可编辑幅度。提出一种基于运动动态性建模的运动自动连接方法。引入隐马尔可夫模型对运动动态性进行建模，通过快速自适应比例高斯过程隐变量模型对运动数据进行降维，将运动数据映射到低维隐空间，在隐空间基于马尔可夫链对运动的时序动态性进行建模，达到了在低维空间分析高维运动数据的目的。为了解决运动连接的两个核心问题，即过渡运动长度估算问题和过渡运动姿势生成问题，提出了基于运动速度的过渡运动长度估算方法和基于隐轨迹插值的过渡运动姿势生成方法。提出一种面向人体四肢的运动重定向方法，包括面向人体下肢的运动重定向和面向人体上肢的运动重定向。对于面向下肢的运动重定向，将运动依据其是否需要对脚与约束面的接触位置进行约束分为两类，针对不需要约束脚与约束面的接触位置的运动提出基于下肢向量特征不变的运动重定向方法，针对需要约束脚与约束面的接触位置的运动提出基于下肢运动轨迹投影变换的运动重定向方法。对于面向上肢的运动重定向，分析了人体上肢在运动过程中受到的约束情况，根据不同约束提出静态约束的运动轨迹重定向方法、动态约束的运动轨迹重定向方法和语义约束的运动轨迹重定向方法，将原始骨骼模型的上肢运动轨迹重定向到目标骨骼模型，以重定向后的运动轨迹作为约束对目标骨骼模型进行逆向运动学求解，最终实现运动重定向。设计并实现了一个基于运动捕获的人体运动生成与编辑原型系统，即三维虚拟人运动生成与编辑系统，对本文所提出的算法和技术进行了验证，系统在相关项目中得到了较好的应用。综上所述，通过本文的研究，解决了基于运动捕获的人体运动生成与编辑中涉及到的若干关键技术，提高了现有运动的可重用性，对运动捕获技术更好地应用于角色运动和计算机动画起到了一定的促进作用。更多还原

【Abstract】 Synthesizing virtual character motion in virtual environment is one of researchfocuses in computer graphics and computer animation, which can be used widely togame, advert, movie, military affairs and so on. As the complexity of virtual characterskeleton, synthesizing realistic motion in computer is very difficult. Synthesizing amotion segment must take animator lots of time and energy even by dint of computeranimation software. As the development of motion capture technology, it became realityto generate virtual character motion with motion capture data. This method is easy toimplement and can generate vivid motion, but existing motion capture data are notenough to satisfy application in practice. So, it is need to reuse existing motion capturedata to generate new motion.The aim of this thesis is to generate vivid human motion, based on existing humanmotion data, by simple interactive control of user. This can reduce the complexity ofhuman motion generation and reuse existing human motion better. The key technologiesinclude: nonlinear dimensionality reduction of human motion, motion generation andediting, motion extension, motion segmentation and join, motion style editing and so on.The thesis is focused on the above associated issues, employing the knowledge fromstatistic learning, pattern recognition, data mining, computer graphics, optimizationtheory, etc.This dissertation focuses on the reconstruction problem of geometries and texturesfrom multiple wide baseline images，some creative algorithms and methods have beenproposed, and the highlighted ideas and main contributions are described as follows:An automatic segmentation method for motion based on detecting motionfeature change is proposed. The motion captured by motion capture equipment ofteninclude several semantic features and the motion segments which include singlesemantic feature are needed in motion production, so the long motion which includeseveral semantic features must be segmented as motion segments. The semantic featureof motion belongs to the category of subjective cognition, the motion is mapped tolow-dimensional feature subspace to avoid the difficulty to analyse and model semanticfeature of motion directly and the motion geometric feature is extracted in the featuresubspace. The geometric feature corresponds to the high-level semantic feature and itschange can denote the change of the high-level semantic feature. This method analysesthe high-level semantic feature by analyzing the geometric feature to implementautomatic segmentation for motion in the semantic level.A method of human motion nonlinear dimensional reduction and generation isproposed, based on fast adaptive scaled Gaussian process latent variable models.Through statistical learning on motion data, the motion data are mapped from high-dimensional observation space to low-dimensional latent space to implementnonlinear dimensional reduction, and probability distributing of posture space whichmeasures the nature of posture is obtained. The posture which meets constraints and hasmaximal probability can be computed as the solution of inverse kinematics. Thismethod can avoid cockamamie computation and posture distortion existing in traditionalinverse kinematics. Then, two methods of motion generation are proposed, which arethe motion generation based on motion trajectory editing and the motion generationbased on key frame editing. Compared with the SGPLVM, the FASGPLVM has higherconvergence velocity and precision and extends editing range of motion by adaptingmotion editing direction.An automatic transition method for motion based on motion dynamic model isproposed. The hidden Markov model is introduced to model the motion dynamics, inwhich the motion data are mapped to low-dimensional latent space by the fast adaptivescaled Gaussian process latent variable models and the motion danymics is modeledbased on Markov chain in the latent space, to analyse the high-dimensional data in thelow-dimensional space. To solve the two key problems, estimation of the length oftransition motion and generation of pose of transition motion, an estimation method ofthe length of transition motion based on motion velocity and a generation method ofpose of transition motion based on latent trajectory interposition are proposed.A motion retargeting method orienting human limbs is proposed, whichincludes motion retargeting orienting human lower limbs and motion retargetingorienting human upper limbs. For motion retargeting orienting human lower limbs,motions are classified to two classes by needing to constrain the contact position of footand constraint surface or not. A motion retargeting method based on lower limbs vectorfeature fixedness is proposed for the motions which do not need to constrain the contactposition of foot and constraint surface, and a motion retargeting method based on lowerlimbs motion trajectory projection is proposed for the motions which need to constrainthe contact position of foot and constraint surface. For motion retargeting orientinghuman upper limbs, the motion trajectory, including static state constraint, dynamicconstraint or semantic constraint, is retargeted to the target skeleton model. Then theretargeted motion trajectory is used as constraint to solve inverse kinematics for targetskeleton model to implement motion retargeting.A prototype system of human motion generation and editing based on motioncapture is designed, named motion generation and editing system for3D virtual human,in which methods of this thesis are validated. Some parts of this system have beenapplied to correlative projects.更多还原