Research On Algorithm Of Vehicle Flow Monitoring And Prediction Based On Deep Learning

With the increase of urbanization population,the number of urban motor vehicles increases year by year,and the traffic problems caused by it have become the urgent problem that the government urgently needs to solve.In recent years,scientists believe that Intelligent Transportation System(ITS)is the most effective method to solve the current urban traffic problems,and traffic flow monitoring and prediction is an important research content in ITS,but the traditional method has low accuracy and slow speed.Therefore,based on the artificial intelligence technology,this paper proposes an optimization algorithm based on the existing deep learning algorithm to achieve the function of traffic flow monitoring and prediction,the main contents are as follows:First,analysis and research of deep learning target detection algorithm.The structure of deep learning convolutional neural network was analyzed,then the principle of Faster R-CNN,YOLOv1,YOLOv2 and YOLOv3 target detection algorithm were studied and analyzed,and finally the vehicle flow data set was made for training.The experimental results show that Faster R-CNN detects a frame of images for about 0.12 s and 0.025 s for a series of YOLO;The mean of YOLOv3 recall rate and average precision exceed 99%,YOLOv2 average is about 92%,the average value of Faster R-CNN is about 90%,and the average of YOLOv1 is only about 72%.Second,YOLOv3 vehicle flow detection algorithm performance optimization.Aiming at the problem of missed detection of YOLOv3 dense vehicles,the Spatial Pyramid Pooling feature enhancement module is used to optimize YOLOv3 to obtain more comprehensive feature information,and the detection algorithms of YOLOv3-P1 and YOLOv3-P3 are obtained.Using the vehicle flow data for training,the experimental results show that the recall rate,accuracy rate,and average accuracy of YOLOv3-P1 and YOLOv3-P3 are increased by nearly 0.1% compared with YOLOv3 under different scenarios and different weather,and the missed detection of YOLOv3-P3 algorithm is more less.Third,research on traffic flow statistical algorithm.Aiming at the problem that a single virtual detection line only counts the number of vehicles in a single lane,a double detection line matching method that can simultaneously count vehicles in multiple lanes is proposed.The double detection line matching method can not only adjusting the width of the statistical area,but also judges whether the vehicle in the current frame has completed the statistics in the previous frame.In addition,the influence of statistical area width and interval frame statistics on vehicle flow monitoring were explored.The experimental results show that the statistical accuracy is greatly affected by the width of the statistical area and the number of interframes.The dual detection line matching method has the highest accuracy of 100% in normal weather and more than 96% in severe weather.Fourth,based on LSTM combined traffic prediction algorithm research.For single deep learning Long Short Term Memory(LSTM)recurrent neural network,machine learning Support Vector Machine(SVM)and Linear Regression(LR)algorithms,it are relatively high that are Root Mean Square Errors(RMSE),Mean Absolute Error(MAE)and Mean Absolute Percentage Error(MAPE)in vehicle flow prediction.First of all,LSTM is used to make preliminary prediction,and then SVM algorithm is used to predict and optimize again to form LSTM + SVM algorithm.The study found that the RMSE,MAE,and MAPE of LSTM + SVM algorithm are reduced by about 7.4%,9.5%,and 9.8% respectively compared with the SVM algorithm,and decreases by about 13.4%,10.3,and 10.1% respectively compared with the LSTM.Then,the LSTM algorithm and the LR algorithm are combined to propose the LSTM + LR combined traffic prediction algorithm.The experimental results show that the combined prediction algorithm reduces the RMSE,MAE,and MAPE by 5%,6.5%,and 7.3% on average compared with LR algorithm,and RMSE,MAE,and MAPE decreases by 11.2%,7.3%,and 6.8%compared with the LSTM algorithmIn summary,the optimized YOLOv3 algorithm can improve the accuracy of vehicle detection,and the proposed double detection line matching method can improve the accuracy of vehicle statistics,and the combination of LSTM algorithm can improve the performance of vehicle flow prediction.The research results provide a new research idea and a complete set of feasible solutions for vehicle flow monitoring and prediction.