基于社会力异常检测改进算法的人群行为模型  

A Crowd Behavior Model based on an Improved Social Force Anomaly Detection Algorithm

作　　者：卢丽 许源平[1] 卢军[1] 黄健[1] 张朝龙[1] 王晶[2] LU Li1, XU Yuan-ping1, LU Jun1, HUANG Jian1, ZHANG Chao-long1, WANG Jing2(1. College of Software Engineering, Chengdu University of Information Technology, Chengdu 610225, China;School of Computing ＆ Engineering, University of Hudderseld, Queensgate, Huddersheld, U)

机构地区：[1]成都信息工程大学软件工程学院,四川成都610225 [2]英国哈德斯菲尔德大学

出　　处：《成都信息工程大学学报》2018年第1期1-7,共7页Journal of Chengdu University of Information Technology

基　　金：国家自然科学基金资助项目(61203172;61202250);四川省科技厅资助项目(2017JY0011;2014GZ0007);深圳重大国际合作资助项目(GJHZ20160301164521358)

摘　　要：社会力异常检测算法(SAFM)是检测人群异常行为(人群聚集和恐慌逃散等)的一种核心算法,提取的底层特征不能完整地描述人群的运动状态,导致人群异常行为的识别率低。为此,提出一种改进的社会力异常检测算法(SFDE)解决此问题。算法引入人群运动的轨迹避免底层特征的丢失,通过无监督方法将轨迹进行聚类,再通过轨迹和人群相互作用力建立人群行为模型(人群运动强度、人群方向熵和聚簇中心距离势能的作用力)。为证明算法有效性,应用改进的SFDE算法结合深度学习模型来识别不同的人群异常行为。通过UMN数据对算法进行验证,结果表明SFDE算法的准确率比传统的SAFM算法提高了18%,并且执行时间提高了2.2 s。Social Force anomaly detection algorithm（ SAFM） is a core algorithm for detecting abnormal crowd behaviors（ e.g.,crowd aggregations and panic escapes,etc.）. Some low-level features of the algorithm can＇t fully describe the movement states of the crowd,so the classification recognition rate is very low.Thus,an improved social force anomaly detection algorithm（ SFDE） is proposed in this research to solve this problem.This algorithm introduces the trajectories to avoid the loss of low-level features,and it also groups trajectories into clusters by applying an unsupervised algorithm.Finally,a model of crowd behavior can be established through combination of trajectories and multiple crowd forces（ e.g.,the kinetic energy of the crowd,the entropy of motion direction and the force of cluster centers）.To test the validity and effectiveness of the proposed algorithm,this paper presents how to apply SFDE together with the deep learning model to recognize various crowd behaviors. The SFDE has been tested and evaluated by using the UMN dataset. Experimental results show that the accuracy of SFDE is 18 % higher than SAFM,and the execution time is decreased by 2.2 s.

关 键 词：社会力 人群异常检测 无监督算法 轨迹聚类 群体行为 深度学习 视频监控 

分 类 号：TP317.4[自动化与计算机技术—计算机软件与理论]