视频序列中运动目标检测与人体关节点的跟踪识别研究

【作者】 周兴龙；

【导师】 金季春；

【作者基本信息】 北京体育大学 ， 人体运动科学， 2006， 博士

【摘要】 计算机视频图像分析在军事、医学、工业、交通等方面有着广泛的应用。把计算机视频图像分析技术应用于体育科学,对运动员的运动过程进行自动识别与跟踪,有助于对运动技术动作进行分析,对运动员科学训练和运动技术诊断等都能起到很好的作用。本论文在查阅大量文献资料的基础,主要对视频图像采集系统的组建、运动目标的检测、人体关节点识别跟踪三方面进行了研究。在视频图像采集系统的组建中,介绍了视频图像采集系统历史和发展,在当前条件下,选取MiniDV数码摄像机和普通PC机作为视频采集硬件。应用Microsoft DirectShow技术,编写了DV视频采集程序。整套视频图像采集系统的软、硬件构成满足运动技术图像分析的要求。在运动目标检测研究方面,根据视频分割算法的主要特点,介绍了基于运动的分割技术、时空分割技术、交互式分割技术。指出了各种不同算法的优缺点。在此基础之上,本论文应用了一种改进的二次帧差运动目标检测算法,通过此算法实验结果,表明本论文的改进的二次帧差运动目标检测算法有较好的检测效果,分割出的图像质量满足快速反馈系统的要求。在关节点位置识别跟踪研究方面,首先介绍了几种常用的目标特征和匹配算法。通过对己有算法的分析,提出了一种基于图像纹理分析的匹配准则,以一块区域的直方图的矩作为像素块的匹配准则。通过改进的四步运动矢量查找算法,进行关节点的匹配识别。实验结果表明此跟踪方法对贴有标志点的关节点识别效果满意,对无标志点的关节匹配效果有待提高。更多还原

【Abstract】 Digital video processing is widely used in military, medicine, industry, transportation and etc. It is useful to apply digital video processing to sports science, by auto-extraction as well as body tracing in image sequence, in scientific training, techniques analysis and diagnosing of sports.Based on studying and organizing the relative materials, this paper involves in the three topics, the establishment of video capture system, the detection of moving object, the extraction and trace of body joint in image sequence.In terms of the establishment of video capture system, after introducing the history and development of the video capture system, we adopted the MiniDV and common PC as hardware of video capture. We programmed a DV video capture program, using Microsoft DirectShow. The constitution of hardware and software of video capture is qualified for the requirement of sports techniques.In terms of the detection of moving object, according to the characteristic of video segmentation method, three segmentation methods are introduced in this paper, based on motion, spatial and interactive. A better algorithm of detecting moving object is put forward, based on twice-frame-difference. Experiments show this algorithm is very valuable and is qualified for the motion quick feedback system. In terms of the extraction and trace of body joint in image sequence, several main object features and matching algorithms are introduced. Base on past algorithms, a new matching criterion relied on texture is posted up. We make the extraction and trace of body joint by an improved algorithm depended on Four-Step Search. The result shows that the tracing method is effective on joints with marker, however, the method needs to be further improved on joints without marker更多还原