摘摘 要要 多媒体技术的高速发展使得用计算机描述和生成高质量的图像和视频等多媒体信息成为可能， 而互联网带宽的提高与计算机处理能力的极大增强促进了多媒体信息的迅速传播。 但随之而来的问题是如何利用计算机理解和再生大量的数字媒体信息。 在这种背景下， 二十世纪九十年代兴起了对数字娱乐的研究与开发。 本文的目的是研究从视频流中提取人体运动，并对提取的运动进行处理，合成满足用户需求的运动的理论、方法和技术，探索利用计算机视觉和图形学再生人体动画的方法。 文中首先讨论了视频流中的人体运动提取方法。针对单视图视频流，我们提出了基于变形块和基于图像差分的特征跟踪方法， 并对算法的性能进行了对比分析；在多视图视频流的运动提取中，我们提出了基于 Kalman 预测和极线约束的特征跟踪、基于多属性量化的特征跟踪、基于 HMM 多属性融合的特征跟踪、不完全运动特征跟踪等算法，并通过对各种算法的性能及应用的对比分析，提出了特征识别可以用来指导特征跟踪的思想。 为了提取精确的三维运动，本文研究了摄像机标定技术、图像的三维重建技术和数据处理方法。 为了再生复杂虚拟场景下逼真的人体动画， 本文对其中的运动合成技术进行了深入研究。在第六章中我们讨论了基于约束的运动编辑方法，该方法主要用来满足场景修改和再生运动的需求，提高运动重用性。在第七章中我们提出了将多个运动融合到同一个非结构化虚拟场景下的多角色运动融合的新概念， 通过角色的自主感知和决策，系统可自动再生出复杂场景下自然逼真的拟人动画。 鉴于难以从影视片段中提取精确的三维运动，我们提出了 2.5 维动画的新概念和实现方法， 通过将从影视视频流中提取的二维运动信息重定向三维动画模型上，可以再生出具有三维效果的逼真动画。 在第九章中，我们介绍了利用提取的三维运动进行三维动画再生的技术，主要包括动画再生方法、 有生命物体的动画模型以及非生命物体的动画模型与人体的结构和运动对应方法、以及基于时空约束的运动重定向等技术。 在第十章中，我们介绍了实现的基于视频的人体动画系统（VBHA） ，里面集成了以上几章中的核心内容，并给出了实验结果。 关键词关键词： 计算机视觉 计算机图形学 视频流 运动分析 特征跟踪 三维重建 运动编辑 运动合成 计算机动画 人体动画 多角色运动融合 Abstract The highly development of multimedia technology makes it feasible that uses computer to represent and produce multimedia information of high quality such as images and videos. Moreover, the improvement of Internet band and the boost of computers processing capability speed up the broadcast of multimedia information. But the problems along with it is how to use computer understand and regenerate enormous digital media. Under this circumstance, there springs up the research and development of digital entertainment. This dissertation aims to research the theory, approach and technology of how to extract human motion from video streams, and then process and synthesize the extracted motion so as to meet the requirement presented by users. And exploit the approach to apply computer vision and graphics to regenerate human animation. In the dissertation, we firstly discuss the approaches to extract human motion from videos. As the single-view video is concerned, a deformed blob based and an image differentiation based feature-tracking algorithms are presented. We also compared the performance of these two algorithms. In multi-view video motion extraction, four algorithms to track features, such as Kalman prediction and epipolar line constraints based feature tracking, feature tracking by quantifying multiple attributes, feature tracking by fusing multiple attributes with HMM and incomplete motion features tracking, are presented. Based on the comparative analysis of the performance and application of such algorithms, we present the idea of applying feature recognition approaches to feature tracking. In order to precisely extract 3D motion, some key technologies, such as camera calibration, 3D reconstruction and 3D data processing, are studied in this dissertation. For regenerating natural human animation adapting to complex virtual scenes, motion synthesis is deeply studied in the dissertation. In chapter 6, we discussed spacetime constraints based motion editing, which is used to modify motion and regenerate new motion so as to improve the reusability of 3D motion. In chapter 7, we presented a new idea of multi-character motion fusion, which fuses multiple motions into one non-structured virtual scene so as to automatically regenerate natural and vivid human animation under complex virtual scene by characters automatic self-sense and self-decision. Considering that it is difficult to precisely extract 3D motion from movies, we presented a new idea of 2.5D animation and its approaches, which maps extracted 2D motion from movies to 3D animated model produced by computer and then produces natural animation with 3D visual effect. In chapter 9, we discussed 3D motion retargeting using extracted 3D motion, the key technologies cover animation regeneration, the correspondence of configuration and motion between human and life-likely animated model or nonlife-likely animated model, space time constraints based motion retargeting and so on. In chapter 10, we introduced Video Based Human Animation (VBHA) system developed by us, which realizes the approaches and ideas above. And we showed some experimental results and analysized them in this chapter. Keywords: Computer vision, Computer Graphics, Video streams, motion analysis, feature tracking, 3D reconstruction, motion editing, motion synthesis, computer animation, human animation, multiple characters motion fusionI 目目 录录 第一章 引言.1 1.1 研究动机.1 1.2 研究问题.2 1.3 论文的组织.3 1.4 本论文的贡献.4 第二章 视频流中的运动提取与运动合成综述.