【作者】 张毅；

【导师】 印勇；

【作者基本信息】 重庆大学 ， 信号与信息处理， 2009， 硕士

【摘要】 人体行为识别指对人体运动的模式进行分析和识别,是计算机视觉领域被广泛关注的研究热点。异常行为识别是人体行为识别的一个重要研究方向,本文主要对跳、跑、呼救、弯腰、挥手五种异常行为进行识别。这项研究在智能安全监控领域有着广泛的应用价值,基于计算机视觉的智能安全监控系统不但能够高效地完成保障安全的任务,还能节省大量的人力物力,给社会带来很大的经济收益。本文主要针对静止背景下的人体异常行为检测与识别技术进行了较深入的研究。本文首先研究了现有的几种主要的运动检测算法:瞬时差分法和背景建模法并分析了这两种检测算法的优劣。在以上分析的基础上,该文将背景减除法与时间差分法相结合解决了传统方法无法从复杂背景中检测到较完整的运动人体轮廓的问题。该方法基于帧间差分法检测出帧中的背景像素点后,再确定每个点的高斯模型,最后运用背景差分准确检测出场景中的运动目标。然后,对视频图像中的阴影问题进行了分析,介绍了在RGB色彩空间和HSV色彩空间中检测阴影的算法。比较在两种色彩空间阴影检测算法的优缺点,改进阴影检测算法,结合两种色彩空间进行阴影的检测,这样在阴影的检测精度和实时性方面都有改善。同时介绍了几种常用的图像后处理技术:中值滤波和数学形态学方法。在运动目标识别方面,以改进的Hu矩不变性算法为基础,针对可见光图像的特点,采用不变矩与图像特征参数组成新的特征向量,对可能的目标进行识别,不仅提高了算法速度,也提高了识别准确率。然后通过模板匹配的方法,采用Hausdorff距离计算所需识别的当前行为特征向量与模板行为(正常行走的行为)特征向量之间的相似性,并通过相应的阈值判定该行为是否为异常行为。对于上述各个算法,本文均进行了模拟实验,结果表明,综合运用本文中提出的图像处理和识别方法,可以较准确、快速地检测和识别人的异常行为。更多还原

【Abstract】 Human action recognition is to analyze and recognize the pattern of human motion, which is a wide interesting research topic in computer vision field. Abnormal behaviors recognize is an important component of human motion analysis and there are five kinds of abnormal behavior, including jump, run, bend, jack, wave in this paper. Research in this domain involves wide application in the intelligent security monitoring system, which not only completes the safeguard tasks efficiently, but also saves a great deal of human labor.Two algorithms of moving object detection are described in this paper, which are the instantaneous difference method and background modeling method. Their advantages and disadvantages are analyzed. An approach of combining the background subtraction and temporal difference is proposed to resolve the problem that current methods can’t obtain the integrated moving human contour. Two consecutive frames are subtracted to pick up background pixel in this method, and then the Gaussian model of every pixel is built. Background subtraction is used to detect moving objects.Next, the shade processing is discussed. The shade detection algorithms in RGB and HSV color space are introduced. An improved algorithm of detecting the shade combining two color spaces is proposed to improve the accuracy and real-time performance. Meanwhile, Image post-processing including median filtering and mathematic morphology method is discussed..An improved Hu moment method is used to abnormal behavior recognition. Hausdorff distance is used to calculate the similarity between recognizing behavior feature vector and normal behavior template one and determine whether the behavior is abnormal or not. Experimental results show that proposed method can detect and recognize human abnormal behavior accurately and efficiently.更多还原