摘要
大多基于深度学习的异常行为识别算法的主要思路是学习单一正常行为带有的特征,并在识别阶段用预测误差或重建误差判断当前帧是否存在异常行为,但由于正常行为的多样性,使得在识别阶段会对正常行为造成误判。为此,本文提出融合记忆模块与编解码框架的异常行为识别算法,通过引入记忆模块来存储多种正常行为的特征。实验表明,与最新的算法相比,本文的算法在公开数据集Avenue和ShanghaiTech Campus上的异常识别AUC分别提升了0.049和0.028。
The main idea of most abnormal behavior recognition algorithms based on deep learning is to learn the features of a single normal behavior and use prediction or reconstruction errors to determine whether there is abnormal behavior in the current frame during the recognition stage.However,due to the diversity of normal behavior,misjudgment of normal behavior may occur during the recognition stage.Therefore,this paper proposes an abnormal behavior recognition algorithm that integrates memory modules and encoder-decoder framework,and stores muliple features of normal behavior by introducing memory modules.Experiments have shown that compared with the latest algorithms,our algorithm has improved anomaly recognition AUC by 0.049 and 0.028 on the public datasets Avenue and ShanghaiTech Campus,respectively.
作者
范冰
顾博涵
FAN Bing;GU Bohan(Institute of System Automation Technology and Application,Guodian Nanjing Automation Co.,Ltd.,Nanjing,Jiangsu 210000,China;Suzhou Medical College,Suzhou University,Suzhou,Jiangsu 215021,China)
出处
《自动化应用》
2023年第22期172-176,179,共6页
Automation Application
关键词
记忆模块
编解码框架
异常行为识别算法
memory module
encoder-decoder framework
abnormal behavior recognition algorithm