基于能量的人群异常行为检测算法研究  

Research on crowd abnormal behavior detection algorithm based on energy

作　　者：李新[1] 郝海江 陈帆 黄琳[1] LI Xin;HAO Haijiang;CHEN Fan;HUANG Lin(School of Information Science and Engineering,Guilin University of Technology,Guilin 541000,China)

机构地区：[1]桂林理工大学信息科学与工程学院,广西桂林541000

出　　处：《现代电子技术》2022年第4期155-160,共6页Modern Electronics Technique

基　　金：广西自然科学基金(2018GXNSFBA281081);广西嵌入式技术与智能系统重点实验室开放基金(RZ18103089,2018A⁃10,2019⁃02⁃03)。

摘　　要：群体异常的发生会危害社会公共安全,因此利用现有技术对人群进行实时监控和分析,对于维护社会秩序和公共安全具有重要意义。针对公共场所中人群异常行为的检测问题,文中提出一种基于人群运动能量变化的方法来检测人群中是否发生异常行为。该方法利用Farneback光流算法获得视频帧的前景运动图像,并计算视频帧的全局光流幅值。由于人群运动能量与光流幅值成正相关,因此可以通过计算全局光流幅值分析人群运动强度,计算人群瞬时能量,通过将相邻帧间的能量差值与特定阈值做比较来判断人群中是否发生异常事件。最后,在UMN数据集上对文中方法进行测试,得出三种场景下的AUC值分别为0.992,0.948和0.978。实验结果表明,文中所提出的算法能够有效地检测出人群的异常行为,可以满足实时性的要求,且具有良好的性能。As the occurrence of group abnormalities can endanger social public security,it is of great significance to use the existing technology to monitor and analyze the population in real time for maintaining social order and public security.In allusion to problem of crowd abnormal behaviors the detection in public places,a method based on changes in crowd kinetic energy is proposed to detect whether abnormal behaviors occur in crowds.In this method,the Farneback optical flow algorithm is used to obtain the foreground motion image of the video frame,and calculate the global optical flow amplitude of the video frame.Because the crowd kinetic energy is positively correlated with the optical flow amplitude,the crowd motion intensity can be analyzed by calculating the global optical flow amplitude,the crowd instantaneous energy can be calculated,and whether an abnormal event has occurred in the crowd can be judged by comparing the energy difference between adjacent frames with a specific threshold.The proposed method was tested on the UMN datasets.The results show that the AUC values in the three scenarios are 0.992,0.948 and 0.978,respectively.The experimental results show that the proposed algorithm can effectively detect the abnormal behaviors in the crowd,meet the real⁃time requirements,and has good performance.

关 键 词：人群异常行为 群体检测 运动能量 智能监控 前景提取 光流幅值计算 异常事件判断 

分 类 号：TN919.23-34[电子电信—通信与信息系统]