【作者】 衣淑凤；

【导师】 沈兰荪； 黄祥林；

【作者基本信息】 北京工业大学 ， 电路与系统， 2004， 硕士

【摘要】 摘要随着经济社会的发展，各种公共场地和设施中的人群流动越来越频繁。如何对公共场合的人群进行有效管理与控制，是不得不考虑的重大问题，智能化人群人数统计方法应运而生。智能化人群人数统计可以用于人群的监测和管理，同时也可用于商业领域如市场调查、交通安全以及建筑设计领域等。它直接或间接地提高了上述场合工作人员的工作效率和建筑设施的利用率，对人群人数统计方法的研究有着深远的意义和广阔的前景。本课题的研究内容是运用图像/视频处理和模式识别等技术对人群人数进行智能化统计。其中，如何有效地提取表征人群个体的特征，如何对人群个体和背景进行分类，以及如何快速统计人群人数是人群人数统计方法的关键技术，也是本论文的主要研究内容。本论文首先介绍了智能化人群监测系统的构成和基本原理以及它的发展。然后指出了现有人群监测系统的不足之处。论文提出了一种通过定位头部识别人群个体从而对人群人数进行统计的方法。首先对头部图像作二维 Haar 小波变换，对一级、二级小波系数进行了分析，并利用反向传播神经网络加以详细验证，最后采用了一级和二级小波的 HL 和 LH 子带系数作为特征量。特征量的选取的理论依据是小波 HL 和 LH 子带能够反映图像的轮廓及横向和纵向的纹理特征。为了进一步降低误检率，我们采用自举的方法对训练样本库进行补充扩大。然后借助基于 YCbCr 空间的肤色模型对实验结果进行了后验证，相比于传统的只考虑色度的肤色模型，我们采用的是最新提出的加入亮度补偿的椭圆肤色模型。最后对结果图像中相同个体的窗口采用目标聚类的分析方法加以合并。实验取得了较好的效果。更多还原

【Abstract】 With the development of the society, more and more people appeared in publicplaces and facilities. So, how to manage the crowd is a problem that we have to payspecial attention to. And intelligent crowd surveillance based on image processingemerges as requested. It can be applied to the fields of crowd surveillance andmanagement, market research, traffic safety, and building design. It may improvework efficiency of the above situations and the rate in use of buildings directly orindirectly. Research on people counting has deep meanings and wide future. The emphasis of the thesis is people counting using technology of image/videoprocessing and pattern recognition. And how to extract the features of head effectively,and how to classify people and background are the key technologies of this thesis. The thesis first introduces the structure and development of intelligent crowdsurveillance system. And then the limitations of present system are addressed. Thisthesis proposes a method of recognizing the individual in the crowd by detecting thehead, and then estimating the number of crowd. After the first and second grade of 2DHaar wavelet transform, we analyses the wavelet coefficients, and useback-propagation network to testify those coefficients selected as features. At last weselect the coefficients of HL sub band and LH sub band as features. The basis ofconcept in selecting those features is that the coefficients reveal the contour and thetexture of horizontal and vertical orientation. In order to reduce the rate of errordetection, Bootstrapping is used to enlarge the number of samples. And as follows askin model of YCbCr color space is taken as post-validation. Unlike the conventionalskin model, a model of adding lighting compensation that is presented newly is used.Finally, the head windows of same person using object clustering are merged. And thetest has got good results.更多还原