Research And Application Of Urban Management Video Surveillance Object Detection Algorithm Based On Deep Learning

In recent years,with the rapid development of artificial intelligence and the joint promotion of academia and industry,a large number of artificial intelligence algorithms have been applied to all walks of life.At present,the new generation of information technology represented by artificial intelligence is accelerating the integration with urban management,providing many new solutions and operation modes for managers to deal with the problems encountered in the process of urban management.The main work of this paper is to combine the computer vision object detection algorithm in deep learning with urban management,and use its powerful learning ability and representation ability to solve some of the pain points and difficult problems.Therefore,this article takes 9 types of illegal objects in urban management video surveillance: mobile vendors,illegal parking of non-motor vehicles,illegal parking of motor vehicles,green damage,illegal setting of billboards,low-lying water,illegal umbrellas,drying along the street,and road damage as the starting point.(1)For some types of problems with a small amount of data,An image data enhancement method based on the improved GrubCut matting algorithm is proposed.By adding a circular LBP operator to GrubCut,the algorithm’s ability to process the detailed texture and position of the cutout is improved.And use the improved algorithm to enhance the data of the two types of samples with a small amount of original data to improve the richness and balance of the training samples.(2)To further improve the problems existing in the one-stage object detection algorithm.On the basis of YOLOv4,DyFPN(Dynamic Feature Pyramid Network)is used to replace the original PANet to improve the problem of increased computational cost caused by the feature fusion stage;at the same time,GHM(Gradient Equalization Mechanism)is introduced to equalize the entire data,Improve the problem of assigning weights to difficult samples and easy samples in classification tasks.Improve the detection ability of the algorithm in the actual scene.(3)Aiming at the problem of application deployment,it is proposed to use Docker as the carrier to embed the optimal model of the trained algorithm into the Docker container,and to develop the software according to the actual business scenario requirements,and release it in the form of a server.Provide urban management staff with a full range of abnormal event perception and intelligent processing capabilities,improve the efficiency of urban management work and the scientificity and reliability of supervision work,and provide strong support for the development and construction of smart cities.