人体运动跟踪的方法与实验研究

【作者】 孙怡；

【导师】 胡家升；

【作者基本信息】 大连理工大学 ， 光学工程， 2002， 博士

【摘要】 近年来，人体运动的跟踪与分析在图像处理与计算机视觉领域引起许多学者的关注。这一课题在智能监视系统、虚拟现实、高级用户接口、运动分析和基于模型的图象编码等方面具有广阔的应用前景。利用图象序列进行人体运动的跟踪与分析包含三个基本内容：(1)从复杂背景中提取运动人体；(2)人体运动的跟踪和标定；(3)人体行为的识别和理解。其中，人体运动的跟踪和标定是人体运动跟踪与分析过程的关键，是进一步识别和理解人体运动行为的基础，人体行为的识别和理解达到了人体运动分析的最高境界，但目前还远达不到这一点，因此研究工作主要集中在前两项上。本文提出了几种基于单目视觉的人体运动跟踪方法，这些方法通过分析图像中躯体呈现的特征，跟踪人体的各个部位，获得人体运动的各种参数，来重建人体运动的过程。 人体运动跟踪系统处理的是由摄像机摄入的视频图像序列。首先经过背景去除，把感兴趣的人体目标从图像中提取出来，然后采用基于模型的和非基于模型的两种方法对人体的局部和全身运动进行了跟踪。论文中，跟踪了两个局部部位腿部和头部的运动，利用人体腿部的几何特征，将原始腿部图像序列变换到相应的距离图像上，然后对人体腿部的四个关节点运动进行了跟踪；根据人类视觉识别特性，采用由粗到细的匹配过程和基于面部的色彩信息对图像中的头部进行了检测，并且在细致匹配中利用了模糊模式匹配方法。在对全身运动跟踪中，根据人体的形状特征描绘了人体区域模型图，对全身步行及体操动作进行了二维跟踪和标定，之后再利用摄像机的透视投影模型、人体关节间的骨骼连接关系和比例参数，把关节点在图像序列中二维坐标恢复为三维坐标，最后用棒状图恢复了人体的三维结构。 论文中给出了利用各种方法处理得到的身体各个部位及全身运动的跟踪结果，并与已有的其他方法作了比较。实验结果表明，本文的基于距离图像的腿部运动跟踪和基于关键帧的人体步行运动跟踪，和以头部为基准将模型与图像序列首帧相匹配的方法具有自己的特色。论文在最后部分，对人体运动跟踪与分析方法所存在的问题及进一步完善提出了自己的见解。更多还原

【Abstract】 Human motion tracking and analysis 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 smart surveillance system, virtual reality, advanced user interface, motion analysis and model-based coding, etc. The procedure of the human motion tracking and analysis from a sequence of images involves three main stages: (1) human body segmentation in a complex scene; (2) human motion tracking and body structure reconstruction; (3) motion analysis and action recognition. As the base of the human action recognition and understanding, human motion tracking and body structure re-construction is the key of the whole procedure. Action recognition is the highest level of motion analysis, but now it is far from application. The researching work mainly focus on the first stage and second stage. This paper proposes several methods of tracking human motion from a single view. The precise correspondence between the human region of the image and the real human body is established through analyzing the feature extracted from the human body region of the image sequence, the motion parameters are obtained, ultimately the sequence of human motion is recovered.The image sequence is acquired by a single camera. Firstly, the human region is extracted from the image by the background subtraction, then model-based and none model-based method are applied to track human motion from body parts to whole body. In this paper, two parts of the body leg and head are tracked. The lower limb is tracked using geometric characteristics, the original image sequence is first transformed to the distance image sequence, and then the four joints of the leg are tracked on the distance image sequence. Based on visual perception, the head is tracked using color information of face and matching is proceeded from coarse to fine matching, fuzzy pattern recognition is also applied during fine matching. In the tracking of whole body, a model is described according to the body shape. Using this model, simple and complicated movements such as walking and throwing motion are labeled, and the 2-D coordinates of joints can be got. At last, the 3-D positions of joints are obtained by the model of perspective projection combined with the relationship among joints, and human motion is recovered by 3-D stick model.The experimental results from body parts movement to whole body movement are given and the tracking methods are compared with existing methods. The experimental results show that the matching method using distance image in leg movement, using key frames in walking tracking, and the matching between the model and first frame based on head location are the contributions to the human movement research. At last the further suggestions for the improvement of the system are discussed.更多还原