Research On Key Technologies Of Traffic Flow Prediction Based On Deep Learning

The development of artificial intelligence promotes the progress of intelligent transportation technology,which promotes the intelligent and efficient traffic management and service,electronic information and safety prevention.The problem of urban traffic congestion is becoming more and more serious in China.In the intelligent transportation system,traffic flow prediction is an important means to solve the problem.The traffic flow forecast with a certain prediction range can provide guidance for the management department,and also provide reference for people’s travel,which is of great significance for improving the urban traffic congestion.Due to the wide application of big data technology,the accuracy of vehicle and license plate detection and recognition has been significantly improved in the video monitoring scene,and the traffic flow data acquisition technology based on license plate recognition arises at the historic moment.In this paper,to solve the problem of low accuracy of irregular license plate recognition,the scheme of multi-task weak supervision is adopted to optimize the license plate image correction method,so as to improve the quality of data acquisition.For traffic flow forecasting,the method of mining similar history and attention is used to alleviate the influence of error propagation in long-term prediction.The research and improvement are carried out around the following aspects:(1)Traffic flow data collection based on license plate recognition.An optimization method of irregular license plate recognition based on spatial transformation network is proposed.In view of the problem of license plate deformation caused by the poor location of camera in the video monitoring scene,an image correction optimization method is proposed.Aiming at the problem that STN network depends on parameter initialization,a multi task weak supervision training method is proposed.In the training,the loss of classification and regression is calculated according to the situation,and a small number of labels are used to improve the positioning ability of the benchmark.This method effectively improves the recognition rate of license plate and is of great significance to collect high-quality traffic flow data.(2)Traffic flow history information extraction based on similarity search.A similar scene sequence search method based on incidence matrix is proposed.The "similar scene" in traffic flow is defined and divided into periodic scene and nearest scene.Aiming at the similarity measurement of traffic flow,a sequence description method based on incidence matrix is proposed to describe the characteristics of traffic flow sequence more accurately.Furthermore,a similar scene search method based on periodicity and nearest neighbor is proposed,which is divided into three channels to search the windows of recent and distant important time points,which can search the important information of historical traffic flow more comprehensively.(3)Long-term traffic flow forecast based on attention mechanism.A hybrid attention mechanism based on time alignment and context is proposed for long-term traffic flow prediction.Apply similar search results to the attention structure.To solve the problem that the traditional attention model can not learn the correct weight through back propagation,a hybrid attention model based on time alignment and context of similar scenes is proposed to make the learning process of the model more targeted.The model pays more attention to the detailed characteristics of time series,better maintains long-distance dependence,reduces the error accumulated with the increase of step size in the prediction process,and has a significant effect on the long-term prediction of traffic flow.