【作者】 冯莲；

【导师】 邹北骥；

【作者基本信息】 湖南大学 ， 通信与信息系统， 2006， 硕士

【摘要】 近年来,人体运动的视觉分析在图像处理与计算机视觉领域引起了广大学者的关注。这一课题在虚拟现实、智能监控、高级用户接口、运动分析、视频压缩等领域具有广阔的应用前景。本文研究基于视频的人体运动跟踪技术,在归纳和总结国内外这一领域的研究现状,对比各种方法的优势和劣势,分析人体运动跟踪的技术难点的基础之上,针对现有的基于模型的人体运动跟踪方法大多需要人工干预,从而不能满足实时性要求这一局限,以及步行运动跟踪研究中双腿自遮挡造成跟踪准确度不高的问题,提出了一种基于单目视觉的人体步行腿部骨架的自动检测和跟踪算法。算法分析人体步行运动的特征,跟踪下肢的五个关节点,获得了步行的各个参数,重建了人体步行运动的过程。主要研究成果如下:1)提出了无标志的腿部骨架自动提取算法。首先将视频分解成许多连续的静态图像帧,经过背景去除,把感兴趣的人体区域提取出来,通过二值化,中值滤波等预处理方法得到只有人体的一个单连通区域,然后用Sobel算子检测出BoundingBox最宽帧中人体下半身的轮廓,根据运动规律及特征找到后腿踝关节点,结合从BoundingBox最窄帧中所获取的腿长依次得到后腿膝关节,跨部关节,前腿踝关节,前腿膝关节四点,从而构建出腿部骨架模型。2)实现了人体步行腿部骨架的跟踪算法。在完成对腿部骨架模型的自动初始化之后,本文对跨关节、膝关节及踝关节分别采用运动建模、圆周相交定点算法、运动预测及预测点周围搜索RGB相似矩形块三种方法确定每一帧中其实际坐标,从而重构出腿部骨架的运动过程。论文以中国科学院自动化研究所下载的步态图像序列库及实验小组自己拍摄的视频为实验素材,用本文提出的算法实现了对步行人体腿部骨架的自动检测及跟踪。实验结果表明,本文算法对骨架的提取及跟踪准确度较高,不仅摆脱了手工标注的约束,还有效解决了双腿自遮挡造成跟踪准确度降低的问题。论文最后给出了全文的总结及对未来工作的展望。更多还原

【Abstract】 Visual analysis of human motion has been receiving increasing attention from researchers in the fields of image processing and computer vision during the past few years.It has a lot of applications in virtual reality,smart surveillance system,advanced user interface,motion analysis and video compressing,etc.This paper focuses on the technology of human motion tracking based on video, First,we make a summarization of the domestic and overseas status of the research in this field.On the basis of this,we analyse the technical difficulties of human motion tracking.As most of the existing model-based methods of human motion tracking perform not so good in some situation as they need mannual intervention, and also the precision of tracking is not so satisfying during the research of tracking of walking people because of the self-occlusion of legs,this paper proposes an algorithm of automatic detection and tracking of legs of the walking people based on monocular image sequences,in which we analyse the features of walking people,track the five joints of lower limbs,get various parameters,and then re-construct the walking process.The main research achievement is as follows:1) We propose an algorithm of markerless automatic extraction of leg skeleton.First we divide the video into continuous image sequences,after background subtraction,the satisfying human region could be extracted,then we get a single-connected region by converting the RGB image to binary image and median filtering.Afterwards,the contour of lower limbs in the frame with a widest BoundingBox is detected,using Sobel operator,to find the ankle joint of leg behind according to the features and rules of walking,then,the joint of knee of leg behind,hip,ankle of leg in front,knee of leg in front could be got in turn.So,model of leg skeleton is constructed.2) We complement the algorithm of tracking of human leg skeleton.After the automatic initialization of the model of leg skeleton,we find out the coordinates of each joint in each frame by motion modeling, pointing by circle interset,motion predection and retangular region matching according to its colour.So the motion of legs could be re-constructed.The experiment materials include two parts,one is the gait database downloaded from CASIA and the other are several videos we shoot ourselves.We achieve on automatic detection and tracking of leg of walking people,using the arithmetic proposed in this paper.The experiment results indicate that the precision of the tracking of leg skeleton is satisfying,using our algorithm.It gets rid of the restriction of mannual intervention,and also the problem of self-occlusion could be solved effectively. A conclusion of the whole text and更多还原