节点文献
视频图像中人体目标的检测方法研究
Algorithm Research for Human Body Detection in Video Image
【作者】 王传旭;
【导师】 刘智深;
【作者基本信息】 中国海洋大学 , 地图学与地理信息系统, 2007, 博士
【摘要】 本文主要研究了基于计算机视觉技术的视频智能监控系统中的关键技术,该系统一般由人体目标的检测、行为理解和高层语义输出三个层次,其中人体目标的检测是后面两个模块的基础,也是本文的选题所在。本文主要在以下三个方面进行了探索和研究,提出了新的改进方法,并通过实验证明了新方法的有效性。主要研究工作如下:(1)视频滤波方面:根据视频图像中脉冲噪声的特点,利用脉冲耦合神经网络(PCNN)的简化模型作为分类器,将图像中的像素分为未被噪声污染的点和污染点两类,对污染点像素再用中值滤波器进行去噪。仿真试验比较,表明本文算法比传统中值滤波算法在能滤除噪声的条件下,还可以更好地保护图像的边界纹理细节。(2)作为人体目检测方法,无论在人体与背景有无相对运动的情况下都应有效。对背景建模的分割方法只能检测到背景中有运动变化的人体目标,在该情况下本文提出了两种改进算法。①针对背景建模中的常见问题,根据混合高斯模型对背景进行建模,可较好适应缓慢变化(光照缓慢变化、摇摆的树叶等)的背景,并在混合高斯模型下对背景区域进行分割,达到检测人体目标的目的。但在光照突变的条件下,该方法检测效果差,本文在此基础上进行了改进,首先提出了一种检测光照突变的简单方法,对光照突变的帧结合像素的空间纹理稳定的特征,进行二次分割。试验结果表明该改进算法提高了传统混合高斯模型分割方法的鲁棒性。②很多算法在检测前景人体目标时,分割出的图像内有较大的空洞,目标图像很不完整,无法用传统的形态滤波方法进行修复。本文提出了一种基于前景人体目标帧内区域邻域相关性和背景区域帧间连续性的分割方法。首先以文献[11]的方法为例进行初分割,文中分析了该方法的原理和缺点,以及检测人体目标出现空洞的原因。作为改进算法,本文利用相关系数计算前景区域某一像素帧内空间邻域像素间的相似度,同时计算其帧间像素间的相关性来进行二次分割,可较好地修复初分割的前景人体区域,得到较为完整的人体图像。(3)前两种方法只能检测到背景中有运动变化的人体目标,有一定的局限性。本文进行了基于人体肤色特征的人体目标检测方法研究。对文献[71]提供的方法进行了分析和新的参数预测方法尝试,该文献试验分析皮肤颜色区域在HSV空间随光照帧间变化是一种布朗运动,即是一种稳态随机过程,因而用维纳一步预测方法进行3-D仿射变换参数的预测。试验结果表明在光照基本恒定时,该方法有很好的灵敏性和检测效果;在一定光照变化范围内,能较好地预测当前帧皮肤区域分布,并能有效检测出人体皮肤区域,对光照变化有一定的鲁棒性。
【Abstract】 The key techniques of intelligent surveillance based on computer vision are researched in this paper. Generally this system consists 3 parts that are human body detection and behavior understanding and high leveled judgment output, among which human body detection is crucial and basis for the latter two modules, that is the point why this paper focuses on human body detection in image sequence. The whole work includes three aspects and there are some innovations, which have been proved valid through emulation experiments. They are introduced as following.(1) video image filter designAccording to the characteristics of pulse noise in video image, a neural network PCNN (Pulse Coupled Neural Network) is applied to aggregate the polluted pixels contaminated by pulse noise and the non-polluted ones, then median filter is used to smooth the contaminated pixels. Tests show this algorithm is more valid in filtering pulse noise and superior in preserving the edge and texture of image.(2) An algorithm to detect human body in video image should work well whether the human ismoving or not. Background modeling segmentation methods are capable to detect human body on the condition that the human body should keep moving, and become void when keep stable. Two improved algorithms are put forward herein.①MOGs (mixture of Gaussians) are used in modeling background, which is apt for slowing changed background (e.g. illuminant slow changes and wavering leaves). But this traditional method becomes void when abrupt illuminant changing. An improved method is put forward to overcome its incapability.②Segmented human body image is not integrated for many method, which could not be sewed up through morphological techniques. A new method is proposed, which is on the basis that the foreground pixels within an adjacent area inside of a human body image are closely correlated; while the interframe background pixels in a fixed position are consistent. Tests prove it works well, which could get more integral human body image.(3)the two above methods are limited to detect human body when human body is always moving in a scene. The following algorithm is based on human skin color detection, which could compensate this limitation. After intensive analysis of method [71], a new data prediction algorithm is adopted. Tests show this algorithm is of high sensitivity and high resolution to detect human skin (e.g. human face) when illuminate keeps stable; and could segment most of skin area when illuminant changes to some extent, which prove robust against illuminant variation.
【Key words】 Intelligent surveillance; Human body detection; Mixture of Gaussians; Bayes classification; Wiener one step prediction;