Vehicle Recognition And Tracking Research Based On Deep Learning Theory

With the increasing importance of transportation,it is urgent to sol ve the problem of health monitoring of in-service bridges,which is low-cost,contactless and does not require closed traffic.In this paper,an automatic vehicle detection and tracking classification system based on video was proposed.The system got the basic types of vehicles,axle load range and real-time trajectory,by deep learning algorithm Faster R-CNN to test the vehicle location,using the kalman filter tracking algorithm for multi-target tracking and data correlation technology,using the VGG-16 convolutional neural network combined with established the vehicle data information for vehicle type recognition.The contents of this paper were as follows.The first step was the establishment of vehicle data and information base.This paper collected and analyzed the regulations on vehicle classification in various countries,and established the vehicle classification standard based on relevant research results.Vehicles were divided into 12 categories according to the corresponding parameters of vehicle appearance and axle load.On the basis of the detailed classification standard,the vehicle database including wheelbase interval,axle load interval and total weight interval was established by collecting the information of automobile announcement.The second step was the training of vehicle type recognition convolutional neural network model.In this paper,two different convolutional neural network models AlexNet and VGG-16 were compared and trained by using the established vehicle database.The training results showed that fine-tuning the pre-trained model on the large open dataset and adopting the data augmentation method could both improve the average accuracy of the two models.At the same time,the accuracy of the VGG-16 model was higher than that of the AlexNet model under the same training parameters.With fine-tuning and data augmentation,the average accuracy of VGG-16 model could reach 98.17%.The third step was the vehicle location detection algorithm.This paper compared the effect of background difference method based on gaussian mixture model and Faster R-CNN in vehicle location detection.The comparison results show ed that although the detection method based on gaussian mixture model was simple and fast,it had insufficient ability to deal with problems in a complex environment.Although a lot of resources were consumed in the early stage,the Faster R-CNN had a stable detection effect and could accurately detect the position of vehicles of different categories in images in various environmentsThe fourth step was the vehicle position tracking algorithm and the vehicle automatic classification detection tracking system design.In this paper,the vehicle position detection results obtained in the third step were used as the initial information of kalman filter target tracking algorithm to create a tracker.Experiments on a video showed that the target tracking algorithm based on kalman filter c ould accurately track all kinds of vehicles in complex scenes.Combined with the content of the previous steps,this paper designed a vehicle automatic classification detection and tracking system based on GUI graphical user interface.