结合模式记忆和自监督注意力的人群异常行为检测方法  

Method of crowd anomaly behavior detection combining pattern memoryand self-supervised attention mechanism

作　　者：宁冬梅[1] 梁莉[2] NING Dong-mei;LIANG Li(College of Computer Information and Engineering,Nanchang Institute of Technology,Nanchang 330029,China;School of Mathematics and Physics,Chengdu University of Technology,Chengdu 610059,China)

机构地区：[1]南昌理工学院计算机信息工程学院,江西南昌330029 [2]成都理工大学数理学院,四川成都610059

出　　处：《计算机工程与设计》2024年第8期2527-2533,共7页Computer Engineering and Design

基　　金：江西省科技厅科学技术研究基金项目(GJJ212111)。

摘　　要：为实现复杂环境下监控视频中异常事件的快速检测和准确定位,提出一种结合正常模式记忆和自监督注意力机制的异常检测框架。记忆机制综合考虑正常模式的多样性和差异性,解决卷积神经网络(CNN)泛化性过强的问题。自监督模块包含遮罩卷积层和通道注意力层,通过遮罩信息预测的自监督训练,提高模型对全局特征结构的理解。公开数据集的实验结果表明,所提方法的曲线下面积(AUC)指标分别达到92.6%和82.7%,性能优于当前其它先进的视频异常检测方法,在轨迹检测标准(TBDC)和区域检测标准(RBDC)指标中表现出优秀的异常跟踪和定位能力。To achieve fast detection and accurate localization of abnormal events in surveillance videos in complex environments,an anomaly detection framework combining normal pattern memory and self-supervised attention mechanism was proposed.The memory mechanism comprehensively considered the diversity and difference of normal patterns,and limited the generalization ability of convolutional neural network(CNN).The self-supervised module consisted of a masked convolutional layer and a channel attention layer,and improved the understanding of global feature hierarchy through self-supervised training of masked information prediction.Experimental results on public datasets show that the area under the curve(AUC)performance of the proposed method reaches 92.6%and 82.7%,respectively,outperforming other state-of-the-art anomaly detection methods,and the trajectory-based detection criterion(TBDC)and region-based detection criterion(RBDC)results validate that the proposed method has excellent anomaly tracking and localization ability.

关 键 词：人群异常行为检测 自监督注意力 卷积神经网络 遮罩卷积 全局特征结构 轨迹检测标准 区域检测标准 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]