Research And Implementation For Checkout Etiquette Standards Based On Video Surveillance

With the rapid development of video surveillance technology and the mass deployments of video surveillance equipment,how to use the massive video data of surveillance systems has gradually attracted attention.Due to technical limitations,the use of surveillance videos in the past was limited to video storage and review.The fast-growing deep learning technology in Computer Vision and significantly increased computing power of hardware have brought possibility to meet the demands which were impossible in the past.This article focuses on the application of surveillance videos at the cash register.In the past,in order to supervise the service standards of cashier staff,managers usually took "unannounced visits" or review surveillance videos,which were time-consuming and laborious,and it was difficult to achieve all-day supervision.To solve this problem and improve the management efficiency and the quality of customer service,a checkout etiquette standards intelligent detection system based on deep learning technology has been developed.The accuracy of event detection algorithm is one of the important indicators for evaluating the practicability of this system.In order to improve the accuracy of the event detection algorithm,the accuracy has been optimized from the perspectives of object detection algorithm,event judgment algorithm and engineering application.The following aspects have been achieved:1.To improve the accuracy of the object detection algorithm at the cash register,two networks are proposed to improve the detection accuracy of small objects and moving object.Based on the attention mechanism,a network optimized for small object detection in the cash register scene is constructed.Firstly,the area where small object may exist is cropped with auxiliary information,then the images are encoded and decoded with the RNN,and finally the object type is output through the classifier.Based on the optical flow channel,an object detection network with additional optical flow information is constructed,and the two using ways of the optical flow channel are explored: the four-channel model and the dual feature extraction network model.Finally,the above two networks are combined to improve the comprehensive performance of object detection at the cash register.2.To improve the accuracy of the events judgment at the cashier registers,two video events detection algorithms based on the multi-branch network structure model are proposed,and comparative experiments are carried out.Using the information obtained through the object detection and posture evaluation algorithms,combined with the prior knowledge of the spatial and temporal logic of the event features,write the event detection module manually.Based on the object detection,posture evaluation,added 3D convolution features of the optical flow and the original image,use the Transformer-encoder and classifier to replace the manual detection module.3.Analyze the business requirements of managers and use web technology to develop an intelligent inspection system for checkout etiquette standards.The system samples the surveillance video of the cash register in real time,and the event detection algorithm analyzes the video and captures corresponding events according to business requirements.Using web development technology through the separation of front and back ends,the real-time detection function module and historical event module are designed and developed.