The Design And Implementation Of The Highway Vehicle-type Classification System Based On Lightweight Convolutional Neural Network

With the development of society and economy,my country’s expressway network is becoming ever more perfect.At the same time,the appearance,load size and usage scenarios of expressway vehicles are constantly changing,and the charging standards are also constantly changing.On January 1,2020,the Ministry of Transport formulated a new version of the "Vehicle classification of the toll for highway ",which is the People’s Republic of China Transportation Industry Standard JT/T 489-2019,which revised the previous vehicle toll standards.If the charging method of trucks is changed,the charging method of vehicle(axle)type shall be adopted.In recent years,with the continuous development of deep learning,the accuracy of convolutional neural networks in recognizing graphics and images has become higher and higher.This paper is mainly aimed at the changes in the new version of toll road vehicle classification standards,and proposes a lightweight convolutional neural network based on the highway vehicle identification system,which is deployed on edge computing equipment,realizes real-time detection of vehicles passing on the highway.It can be utilized to highway vehicle charging,ETC vehicle information matching and vehicle inspection.The highway model data set used in this paper,which is divided according to the new version of highway vehicle charging standards,including three categories:passenger cars,trucks and special operation vehicles,and is subdivided into 18 categories labels,including images of multiple highways and various weather environments.This paper adopts the method of data enhancement to expand the data,including cropping,translation and mirroring in the geometric transformation method,and adding noise and changing the brightness in the color transformation method.The data set is comprised of 19302 images.Based on the target detection algorithms YOLOv3 and Mobilnetv3_large,this paper designs a lightweight convolutional neural network model,constructs three network output layers and builds the DBR module,namely two-dimensional convolutional layer,batch normalization layer and Relu6 activation Function layer to improve the running speed of the model.Use GIoU loss function,Soft-NMS method and K-means clustering algorithm to enhance the detection accuracy of neural network models,and build an experimental platform to design comparative experiments,and improve the efficiency of model experiments through parallel training.After testing,the paper proposes a lightweight the mAP value of the type convolutional neural network model is 90.72%,which can effectively detect the type information of vehicles passing on the highway.In this paper,the evaluated lightweight convolutional neural network model is deployed to the edge computing device NVIDIA Jetson TX1,which can be installed on the highway platform to distinguish the traditional video data transmission to the background server for processing and identification,which can be completed Real-time detection of vehicles on the highway saves data transmission time and cost.Using the TensorRT framework to optimize the neural network model,integrate the convolutional layer,batch normalization layer and Relu6 activation function layer,and realize the deployment and inference operation of the convolutional neural network model in hardware devices by merging the splicing layer and other operations.The average time of detecting a frame of image is 35.89ms,which realizes the real-time detection of the type information of vehicles passing on the highway.