多种人群密度场景下的人群计数  

Counting people in various crowed density scenes using support vector regression

作　　者：覃勋辉[1] 王修飞[1] 周曦[1] 刘艳飞[1] 李远钱[1] 

机构地区：[1]中国科学院重庆绿色智能技术研究院,重庆401122

出　　处：《中国图象图形学报》2013年第4期392-398,共7页Journal of Image and Graphics

基　　金：重庆市重大科技攻关项目(cstc2011ggC40009)

摘　　要：公共场合中采用摄像机实现人群计数在智能安防领域具有重要价值,但摄像机透视效果、图像背景、行人相互遮挡等因素制约着人群计数研究的发展和应用。提出一种采用回归模型估计人数的算法。首先,为了消去摄像机透视对图像特征的影响,用图像中行人身高作为尺度基准将图像分成多个子图像块。其次,采用simile分类器优化子图像块的先进局部二值模式(ALBP)纹理特征,并根据子图像块的人群密度,采用两种核函数的支持向量回归机(SVR)建立输入特征和子图像块人数的关系。最后,相加所有子图像块人数得出图像人数。实验结果表明,本文算法测试稀疏人群的绝对误差约为1人,测试拥挤人群的相对误差小于10%,是一种准确率高适用性强的人群计数算法。The use of video surveillance in for people counting public places has an important value in the field of intelli- gent security. However, there are several factors such as camera perspective, background clutter, and occlusions, which restrict its development and application of the study. An algorithm based on the regression model is proposed for estimating the number of people. First, in order to eliminate the effect of the camera perspective on the image features, the input image is divided into several sub-image blocks according to the change of pedestrian height in the image. Second, the simi- le classifier is used to improve the advanced local binary patterns （ALBP） texture feature of the blocks. Then, according to the crowd density, we use the support vector regression （SVR） , which has two kernel functions to establish the relationship between input features and the number of people. Finally, adding the number of persons of all sub-image blocks gives us the total number of people on the image. Experimental results show that the absolute error of the sparse population is approximately one person using the presented algorithm and the relative error of the testing crowded population is less than 10%. This therefore demonstrates the high accuracy of this algorithm, which can be applied for people counting in video surveillance.

关 键 词：人群计数 simile分类器 支持向量回归机 人群密度估计 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]