复杂场景下实时监控中人群密度估计的研究与实现

发布时间：2018-04-09 19:26

本文选题：人群密度估计　切入点：非参数背景模型　出处：《电子科技大学》2011年硕士论文

【摘要】：随着城市化建设进程的加快以及经济社会的高速发展,如娱乐活动、展览活动、体育赛事、庆祝活动等这些大规模的人群密集活动将会愈加频繁出现,这些公共活动的安全问题成为相关部门关注的焦点,同时在计算机视觉和数字图像处理领域安全监控中人群密度估计方法也成为研究的热点。本文研究的重点在于室外实时监控中人群密度估计方法。通过分析国内外对人群密度估计研究现状和所取得的进展,采用基于像素特征和纹理特征相结合的方式分别对低密度人群和高密度人群进行人群密度估计。通过自适应背景建模得到背景模型然后利用图像分割得到人群前景,在对前景人群进行阴影抑制和形态学处理的去噪操作后计算前景面积,若其占整个图像面积比例较大时将其判定为高密度人群图像,反之视为低密度图像。在低密度情况下,对提取出的人群前景进行轮廓检测并计算轮廓像素数目,最后根据多个低密度人群图像样本的计算结果进行最小二乘直线拟合得到人群数目关于前景轮廓像素数目的直线方程,此后就可以利用该直线方程通过计算人群前景轮廓像素数目大概估计出人群数目。在高密度情况下,将高密度人群分为高、偏高和极高三个密度等级,利用灰度共生矩阵进行纹理分析,提取常用的纹理特征,然后通过主成分分析法确定最重要的4个特征作为高密度人群图像的纹理特征,然后采用支持向量机进行训练分类。本文采用了一种基于影响因素描述的非参数背景模型,实验证明用其在室外复杂场景中建模得到的背景图像十分清晰,具备较强的鲁棒性;此外,将人群前景二值化操作融入前景提取过程,轮廓检测采用形态学处理的思想以及特征提取利用主成分分析法等对实时性有极大的提高。
[Abstract]:With the rapid development of city construction and to accelerate the process of economic and social activities, such as entertainment, exhibitions, sports events, celebrations etc. these large-scale crowd activities will be more frequent, security issues of these public activities become the focus of the related departments, both in computer vision and digital image processing method has become a research the hot field of safety monitoring of crowd density estimation.
This paper focuses on the crowd density estimation method of outdoor real-time monitoring. Through the analysis of domestic and foreign research status and the progress of crowd density estimation, using pixel features and texture features based on the combination of low density and high density population population into the crowd density estimation. Through adaptive background modeling and background model the use of image segmentation to get people in prospect, prospects for the crowd to suppress shadow and morphological processing denoising operation after calculating foreground area, if it occupies the entire image area when it is judged as a larger proportion of high density population and low density images, as images.
In the low density case, contour detection of foreground extracted crowd and calculate the contour pixel number, according to the calculation of multiple low density population image sample results are get linear equation groups regarding the number of foreground contour pixels the least squares fitting, then it can be calculated through the crowd foreground contour pixel number roughly estimate the number of people by using the linear equation.
In the condition of high density, high density will be divided into groups of high, high and very high density level, using gray level co-occurrence matrix for texture analysis, texture feature extraction commonly used, and then determine the 4 most important features of the image texture features as high density population through principal component analysis, and then the classification of training support vector machine.
This paper uses a nonparametric background model is described based on the influencing factors, the experiment shows that the background image obtained in the outdoor complex scene modeling is very clear, with strong robustness; in addition, the crowd binarization operation into the foreground extraction process, contour detection using morphological processing and feature extraction using the principal thought component analysis is greatly improved in real time.

【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2011
【分类号】：TP391.41

【引证文献】