Research On Personnel Legitimacy Detection System Based On Deep Learning

With the rapid development of information technology,people’s demand for personnel legitimacy detection in the field of security is becoming more and more intense.At present,personnel legitimacy detection is mainly screening from a large number of monitoring data,which requires manual review,a large amount of manpower and material resources,and is easy to miss detection.Therefore,using deep learning to assist traditional manual video surveillance analysis has become a research hotspot in the field of security.From the perspective of personnel legitimacy detection,this thesis carried out the following work in the two fields of mask wearing detection and personnel abnormal behavior detection:(1)In the context of COVID-19,in order to prevent crowd gathering from causing the spread risk of the epidemic,this thesis proposes a mask wearing detection algorithm based on improved YOLOv4.Aiming at the problem that the detection target is small and the loss function gradient appears to disappear in the process of model training,this thesis uses dense connection structure to improve the Res Block of YOLOv4 backbone network.The feature extraction ability of the model is strengthened,the detection accuracy of small targets is improved,and the gradient is prevented from disappearing.In view of the problem that CIOU loss function does not solve the difference direction between the real box and the prediction box,this thesis uses SIOU loss function to replace CIOU,which solves the difference direction problem and can accelerate the training convergence of the model.In order to solve the problem of uneven sample types,Focal loss function is used in this thesis to make the model pay full attention to the hard-to-classify samples during training(2)Aiming at the problem that the current abnormal behavior detection model cannot guarantee the real-time performance and has low detection accuracy in complex scenes,this thesis proposes an abnormal behavior detection algorithm based on YOLOV5.Firstly,Swin Transformer network is used to replace YOLOv5 backbone network to solve the accuracy problem in complex scenarios.Secondly,aiming at the problem that the model is not sufficient to extract features at different scales and cannot be effectively utilized,this thesis uses multi-scale deeply separable convolutional network to improve CSPLayer,which reduces the number of parameters to a certain extent and ensures that features can be integrated at different scales.The CA attention mechanism is embedded after the Upsample module to enhance the sensitivity of the model to information such as direction and position,which can effectively improve the performance of model checking.8(3)This thesis studies and designs a personnel legitimacy detection system based on deep learning.Based on the analysis of user and functional requirements,the system structure and functions of the module are designed,and the detection algorithms of wearing-mouth clutter and personnel abnormal behavior are deployed into the system.Finally,the detection performance of the system is verified by testing.