Research On Intelligent Car Following Based On Machine Learning

Confronting heavy traffic on urban roads,following vehicles can’t overtake the lead vehicles by changing lanes,which seriously obstructs the smooth passage of vehicles.Therefore,research on car-following behavior has become the key to solving traffic congestion and ensuring traffic safety.It has been a hot topic in the research of intelligent transportation system to reveal vehicle driving characteristics by referring to a mass of vehicle tracking data and comprehensively considering the interactive information among drivers,roads and vehicles.And it’s of great significance to fit a high-precision car-following model for realizing smooth traffic and high safety under the circumstance of unmanned driving by analyzing the real vehicle driving data and applying machine learning theory.To this end,car-following models based on either LSTM neural network(LSTMNN)or Bi-LSTM neural network(Bi-LSTMNN)have been established by combining the theoretical car-following model and considering the timing of vehicle driving data.The neural networks need to be trained by the processed NGSIM data set.The prediction accuracy of the trained neural networks needs to be compared with that of untrained ones by taking comfort and safety as the evaluation criteria for the car-following models,and then the Bi-LSTM car-following model is supposed to be optimized.The main contents and innovations are as follows:(1)Architectural design of car-following model based on(LSTMNN).In the study of car-following process,it’s difficult to depict complex car-following behaviors by a model of car-following driven by a single theoretical factor,and the single-data-driven car-following model lacks clear physical significance.To address this need,based on the GM theoretical car-following model,FV speed,LV speed and the spacing between FV and LV are used as the input while the FV acceleration as the output for the LSTMNN car-following model.The final goal is to establish an LSTMNN car-following model with three hidden layers of LSTM units and the memory effect of drivers.(2)Data processing method and verification of car-following model training.In order to improve the reliability and accuracy of the current NGSIM data set in model training,the smooth function in Matlab software is used to smooth the data and denoise the raw data.Then,to ensure the consistency of car-following behavior,the car-following data of small vehicles on single lane with car-following time over 45 s are selected,and the car-following characteristics are verified.Finally,the car-following data needs to be further filtered and normalized by setting the driver reaction time as 1s.(3)Training and evaluation of car-following model based on LSTMNN.Through the theoretical analysis of the error transfer process of LSTMNN training algorithm,combined with the characteristics of car-following behavior,sigmoid is determined as the activation function,and the learning rate,batch size and maximum iteration times are selected and set as 0.001,128 and 2000 respectively;in the light of the MSE loss function,the LSTM car-following model is trained with Adam function as the optimizer.In addition,a BP neural network car-following model is designed with the same optimizer and super parameters.Taking comfort and safety as evaluation criteria,the comparison results of the two neural networks show that LSTM car-following prediction model has higher reliability.(4)Establishment and optimization of Bi-LSTM car-following model.Car-following behaviors have forward and reverse time dependency,in other words,current car-following state will be affected by the preorder and subsequent states.Considering this,the Bi-LSTM car-following prediction model is proposed.Its prediction accuracy can be improved by adding the reverse learning of the data sequence and fully training the car-following data.Through the training and simulation comparison of the designed model architecture and parameters,it is concluded that Bi-LSTM car-following model is better than LSTM car-following model which only considers one-way data dependency.In order to further improve the prediction accuracy of Bi-LSTM car-following model,optimizer selection and super parameter optimization have been done for the designed Bi-LSTM car-following model.A conclusion has been drawn that its prediction accuracy reaches the highest when Adam function is used and the number of hidden layers,hidden layer nodes and truncation steps is respectively set as 4,16 and 8.