Design And Implementation Real-time Vehicle Information Recogntion System Based On Deep Learning

With the acceleration of the economic development of modern society,smart cities,transportation and driving have become social development trends.The current deep integration of smart transportation and emerging technologies such as artificial intelligence,the Internet,big data,cloud computing,and next-generation communications has brought many new opportunities and challenges.With the increasing number of vehicles,the collection and management of vehicle information has become an indispensable part of smart transportation.The current society’s management of vehicle information is mostly based on plates,and it is impossible to judge the behaviors of unlicensed vehicles and fraudulent plates,which can no longer meet the growing demand for vehicle information management.In the current society’s huge demand for vehicle information management,this paper designs a system for real-time detection of vehicle information based on the entrance of the parking lot.It detects vehicle logos,types and plate characters,and compares with the vehicle data in the established database to determine whether the data are completely matched.When they are completely matched,it is allowed to pass,and when at least one item of data does not match,it is judged as an unknown vehicle and not allowed to pass.The vehicle information detection algorithm of the system in this paper is based on two deep learning network models.The vehicle model and vehicle logo detection,plate positioning are based on the improved YOLOv4 network.The K-means++ algorithm is used to replace the K-means algorithm in the original network,and to reduce the amount of data,speed up network training,introduce separable convolution to replace part of the convolution operation in the original network.Plate recognition uses an improved CNN+GRU+CTC network for character recognition,and improves the loss function to improve the character recognition rate.The data set used by the vehicle logos and types recognition network are produced by the author using label Img software.The pictures are derived from the open source vehicle data set of the University of Science and Technology of China and the author’s photos taken by the author for uniform classification.Finally,a database containing 4 vehicle types and 10 vehicle logos with a quantity of 2814 was produced.The plate recognition network training uses a plate data set generated based on the confrontation network.The data set has been subjected to operations such as tilting,blurring,and adding stains,and has strong robustness.Through testing,compared with the traditional vehicle detection system that only detects the single feature of plates,this system uses the algorithm of the deep learning network on the basis of multi-feature fusion.The system not only realizes the simultaneous recognition of the three characteristics of the vehicle logos,types and plates,but also has a high recognition rate.Therefore,this system has the characteristics of multiple features and high recognition rates.