| Over-limit transportation is a predatory use of highways.It is the main factor leading to the sharp decrease in the service life of bridges and highways and Traffic accident.Informatization is an effective means to control Over-limit transportation.It is urgent to fully realize the automation and informatization of road surface monitoring,provide real-time data for scientific management of roads,overrun and overload management,and provide law enforcement decision-making basis for integrated overrun control.License plate recognition is the cornerstone of the Intelligent Transportation Systems and an important part of the highway no stop Over-limit transportation detection system.It is of great significance for building the Intelligent Transportation Systems and realizing a smart city.However,traditional license plate recognition methods are easily affected by the weather,light,dust and other complex environments,resulting in reduced accuracy of license plate recognition and limiting the application of license plate recognition in real-time traffic.In this paper,artificial intelligence method is used to recognize license plates,and the task of license plate recognition is divided into two parts: license plate location and character recognition.The simplified YOLOv3-Tiny network is used for license plate location,and the convolutional neural network is used to process character recognition,which avoiding cumbersome processing steps.A license plate recognition algorithm with powerful real-time,high accuracy,robustness and easy deployment is proposed,which can realize real-time recognition of license plates under complex conditions,and can be applied to scenarios such as autonomous driving and intelligent transportation,.The main work of this paper is summarized as follows:1)This paper constructs a license plate recognition method based on the YOLOv3-Tiny simplified model.Test on the validation set(223329 image samples),m AP＿0.5 reached 99.5%and F1 reached 99.8%.It has a very good positioning effect on rainy days,nights,uneven illumination,large angles of tilt,and multiple license plate detection.Using a single RTX2080 Ti,the average inference time is only 1.5 milliseconds,the real-time performance is very good.2)This article proposes a very concise network architecture for the license plate character recognition,adopts an end-to-end model design scheme,avoids complicated preprocessing work,simplifies the processing process,and used real license plate data and artificially generated license plate data for training and testing.The model has a recognition rate of 98.63% for conventional license plates,and it also has a good performance in complex situations.According to the model performance of artificially generated license plates as a data set,the model also has a good recognition effect for Chinese characters.3)This article designs an ARM embedded hardware platform based on the i.MX6 ULL processor.The Linux system was built on the hardware platform,and the NCNN framework was transplanted and deployed.The license plate location recognition model was completely deployed in the embedded platform,and the performance of the model was tested and evaluated under that platform.The full set of hardware,software,model training,and model deployment procedures for deep learning algorithms presented in this article have a strong reference significance for the actual deployment of deep learning algorithms. |
