| With the development of a powerful transportation country,in order to improve the service level of roads,the state adjusted the charging mode of highway freight cars in 2020,and changed from the original weight charging to the charging according to the truck model.There are two methods for traditional vehicle identification of Highway Freight Cars: one is to embed portable tape switch or piezoelectric cable in the road surface layer to count the number of axles,which needs to break the pavement layer installation and maintenance,which seriously affects the traffic efficiency;the other is to conduct axle number statistics through high-precision grating sensor and laser sensor detection,which is greatly affected by dust and water accumulation.With the development of deep learning in the direction of target detection,this paper proposes to identify the truck and its axle by using the target detection method,and then identify the vehicle model of the truck.This method makes use of the existing road camera resources to facilitate the implementation and deployment,and has a high research value.The specific research contents are as follows:(1)Two methods are used to collect the side image of the vehicle and establish the data set of the side image of the truck as the research basis.One is to use the existing roadside camera resources to take the truck image on different roads,the other is to crawl the side image of the truck from the network to meet the multi-scale characteristics of the image input in the practical application,and improve the robustness of the network model.In view of the problem that some images in the data set are affected by the light weather,the image is preprocessed based on Retinex theory and the visual enhancement method such as CLAHE,and 164303 available images are obtained.Then,the image is labeled by the annotation tool,and the output is Pascal VOC format,which meets the requirements of the number and format of the target recognition data set.(2)The application methods of deep learning in image field are studied,including convolutional neural network,single recognition network YOLO and Faster rcnn.Through theoretical analysis,YOLO unifies classification and regression as regression problem,reduces model parameters,accelerates forward propagation speed and helps to improve detection speed.For the purpose of experimental comparison,the experimental results show that Faster rcnn is more accurate than YOLOv3 and the position information of prediction frame is more accurate,but the detection speed is low,which can not meet the real-time requirements.Therefore,this paper chooses YOLO3 as the framework of vehicle identification network and further studies on YOLOv3 optimization.(3)This paper optimizes the YOLOv3 network structure from three aspects to improve its detection accuracy.Firstly,the size of the prior frame of data set is obtained by K-means clustering algorithm,which makes the prediction frame fit the target better.Secondly,the resolution of data sets is analyzed.Because the image resolution of the data set is high,some details will be lost when inputting the network,so the network input is increased.Finally,the attention module between channels is introduced to optimize the network model,which improves the weight of the channel with target characteristics,and then improves the accuracy of target detection.(4)The membership relationship between the axle and the truck is studied under the condition of multi truck in single frame image.The position information of the center point of the axle and the prediction frame of the truck is proposed to determine.Through experiments,this method can solve the problem of truck model judgment.In addition,the vehicle identification system is designed and developed in combination with the specific application scenarios.The method of vehicle identification based on deep learning has achieved good results in the self built data set,which basically meets the requirements of real-time and accuracy. |
PDF Full Text Request