人体运动捕捉涉及计算机视觉、图像处理、模式识别和软硬件开发制造等热门领域，有着广泛的应用前景，如安全领域的智能监控、人体运动分析、虚拟现实下的军事训练、人机交互的远程教育、娱乐游戏和电影制作等。本文结合运动捕捉和人机交互，通过单视频捕捉数据来驱动虚拟人模型，分为五步：背景分析和模型初始化，运动检测，人体跟踪，视觉理解以及虚拟人交互。首先，提出一套运动序列，确定了目标的身高和臂展等参数和比例，并根据身体表面颜色进行建模，提取特征，为人体跟踪提供模板。然后，深入研究三种运动检测方法，讨论了基于模型、区域、轮廓和特征的四种跟踪方法。接着，通过特征模板的匹配来检测人体各部分，合理选取关节点，通过缩放正投影原理逆向获取三维运动数据。最后，选取三维棍棒人体模型，结合MoCap代码完成与虚拟人模型的交互。本文的主要研究和工作包括：1. 研究了人体运动分析的流程和方法，讨论了虚拟人驱动的建模问题。2. 提出一套新颖简单的运动序列，可以有效获取人体骨架参数，并通过人体表面颜色的建模，为运动跟踪提取特征模板。3. 改进了背景减除方法，引入关键帧概念达到背景模型的自动更新。4. 搭建了人体运动捕捉系统，通过特征模板匹配来实现人体跟踪。关键字：人体运动跟踪 人机交互 计算机视觉 虚拟人模型

Human motion capture, is a very comprehensive field, related to image processing, computer vision, pattern recognition, software development and hardware manufacturing, with potential applications in smart surveillance, motion analysis, virtual military training, remote education, encoding region of interests, amusement product, etc.This paper combines research of human motion capture and technology of human computer interface to drive virtual human in 3-dimensioanl space by human motion data from single-video capture. From processing aspect, the whole research is divided into 4 steps, and they are human model initialization and background analysis, motion detection and human tracking, pose estimation and behavior recovery, and driving virtual human model.Firstly, a novel and easy-to-apply motion serial is proposed from which human model parameters can be determined, human figure length, torso width and leg’s length, and these invariants are extracted as templates for future tracking.Secondly, 3 kinds of motion detection methods and 4 kinds of human tracking method are discussed and tested. Motion detection methods are mainly background subtraction, temporal difference and optical flow, while human tracking methods include model-based, region-based, contour-based and feature-based. And then a new background subtraction method is promoted which can dynamically adjust itself to environment changes.After motion detection and human tracking, down-to-up approach is employed. Firstly, human body parts are detected and estimated, and then key joints’ parameters are reasonably determined by joint-pair’s constraint. So 3-dimensional human motion data are obtained by inverse projection and saved in BVH format.In the end, we choose 3-dimension figure human model, and combine Blender and MoCap software to show the result. The simple human model, with precise software, future promote the effectiveness, redundancy reducing and easy initialization.In this paper, the main works include:1. Research workflow and methods in human motion analysis, and discuss problems in driving virtual human model;2. Propose a easy-to-apply human motion serial, from which 3D human model is well initialized;3. Promote background subtraction method in motion detection, and develop a dynamical background modeling method which can adjust to the time varying and update with automatic learning;4. Simulate virtual human motion driven by single-video human motion capture data. Combined with Blender and MoCap software, results show that the data is precise and effectively drives the virtual human stick figure model.