介观小团体运动聚类的人群异常检测  

Abnormal crowd behavior detection based on motion clustering of mesoscopic group

作　　者：张旭光[1] 王梦伟 左佳倩 李小俚[1] 

机构地区：[1]燕山大学电气工程学院,秦皇岛066004

出　　处：《仪器仪表学报》2015年第5期1106-1114,共9页Chinese Journal of Scientific Instrument

基　　金：国家自然科学基金(61271409);国家杰出青年科学基金(61025019);中国博士后科学基金(2012M510768;2013T60264);河北省自然科学基金(F2013203364);中国留学基金(2011813018)项目资助

摘　　要：传统的人群行为分析主要基于宏观尺度或微观尺度,前者忽视了个体运动的信息,难以定位局部异常,后者在人数众多的场合下难以精确检测与跟踪行人个体。为了在人群的宏观与微观特性之间搭建一个桥梁,从介观尺度分析入手,提出一种基于人群介观小团体运动聚类的异常行为检测方法。首先通过计算光流场并对其进行平滑滤波来获得人群运动的微观运动描述;进而通过Mean Shift算法对微观局部运动进行聚类,并根据速度、位置信息将微观类别合并为介观人群小团体;最后将速度场特征空间划分为正常和异常区域,根据聚类中心在该空间的区域归属及在该区域的持续时间长短来检测异常行为,并将异常类别所有样本点映射到图像坐标系,采用重心法确定异常发生的具体位置。实验结果表明,提出的方法不但能检测人群运动的全局异常,而且能定位局部异常行为的发生位置。其中,全局异常行为的检测总正确率达99.23%,局部异常的检测总正确率可达86.25%。Traditional crowd behavior analysis method is usually based on macro or micro scale analysis. The former usually fails to locate the local abnormal behavior because the individual motion information is ignored,while the latter usually fails to detect and track individuals accurately in large-scale crowd. In order to build a bridge between the macroscopic and microscopic characteristics of the crowd,this paper focuses on the mesoscopic scale to analyze the crowd behavior,and proposes a crowd abnormal behavior detection method based on the motion clustering of mesoscopic pedestrian groups. Firstly,the optical flow field is calculated and smoothed to obtain the micro description of crowd motion. Then the Mean Shift algorithm is used to cluster the micro local motion; and meanwhile,the micro categories are merged into mesoscopic crowd group according to the velocity and location information. Finally,the velocity field feature space is divided into normal area and abnormal area; the abnormal behavior is detected according to if the cluster center falls in the abnormal area or not and the time the cluster center sustains in the abnormal area. All the sample points of the abnormal class are mapped from velocity space to image coordinate system,and the gravity method is adopted to obtain the location where the abnormal behavior happens. The experiment results show that the proposed method can not only detect the global abnormal crowd behavior,but also locate the position where the local abnormal behavior happens. The total correction rates reach 99. 23% for detecting global abnormal behavior and 86. 25% for detecting local abnormal behavior.

关 键 词：人群分析 异常检测 介观运动分割 Mean SHIFT 

分 类 号：TH744[机械工程—光学工程]