Research On The Key Technology Of Human Abnormal Behavior Detection Under Monitoring Scene

In recent years,with the development of economy and society,people are exposed to a more complex living and working environment,which leads to a higher requirement for security monitoring.As one of the most important part of the intelligent security monitoring system,abnormal behavior detection technology has attracted much attention from many researchers.Abnormal behavior detection system can work all the time,which help to improve the monitoring efficiency,save labor costs,and effectively prevent the occurrence of dangerous events.This thesis studies the key technologies of abnormal behavior detection in security monitoring scene,including the representation and modeling for behavior,and outlier detection technology in unsupervised abnormal behavior detection.1)Representation and modeling for behavior in unsupervised abnormal behavior detection: In order to tackle the problem of insufficient feature learning caused by identity mapping in the existing deep generative model based methods,a two-stream spatial-temporal abnormal behavior detection algorithm is proposed,which consists of a appearance prediction network and a flow generation network,and is able to avoiding the risk of identity mapping and fully exploits spatial and temporal information.In addition,a self-attention module is introduced so that important regions can be paid more attention to.It is proved in experiments that the two-stream spatial-temporal network performs well in identity mapping suppression,and the important regions are also successfully paid attention to by self-attention module.2)Outlier detection technology in unsupervised abnormal behavior detection: This thesis points out the "the phenomenon of uncontrolled of boundary learning" in deep generative model based methods unsupervised abnormal behavior detection methods,then analyzes that this phenomenon is mainly caused by the insufficiency and bias of training data,and the generalization of neural network.Therefore,we propose a general solution by introducing adversarial samples and adversarial training,and applies it into spatial-temporal self-attention anomaly detection network.It is shown in experiments that adversarial samples and adversarial training strategy effectively remit "the phenomenon of uncontrolled of boundary learning",and improve the performance of anomaly detection.This thesis finds and points out two problems in current abnormal behavior detection methods,and attempt to solve them by proposing adversarial training based spatial-temporal self-attention network,which improves the performance and provides a further reference to the future study.