Analysis Of Driving Behavior Based On Deep Car-following Network

Car-following and lane-changing behavior is an important factor that affects the safety of vehicles,and it is also a research hotspot in the field of driving behavior and active safety of vehicles.The study of car-following behavior puts forward stimulation-response,safe distance and other mathematical models to reveal the driving behavior between car-following vehicles on one-way streets,the car-following model can not really reveal the actual driving characteristics of the driver under a few stimulus sources,so it can not be widely popularized.Based on the research status and background of Longitudinal car-following behavior among vehicles,this paper designs a self-learning deep car-following network to explore the driving rules of car-following and lead car in traffic flow,finally,the driving behavior of the driver is reflected by the changing trend of the three parameters:the lead vehicle speed,the rear guide vehicle speed and the relative distance.The main research contents are as follows:(1)Lead vehicle detection based on deep car-following network.In order to ensure the consistency of data distribution in forward and backward propagation,a deep car-following network target detection algorithm is designed based on the principles of CNN convolution network and YOLO series image detection.in order to avoid the over-fitting problem of the network in the training process,the basic components of the network are the batch normalization algorithm and the convolution layer.the results show that the detection performance of YOLO and deep car-following network is better than YOLO in different light environments.The proposed method improves the accuracy of lead car detection.(2)Analysis of driving behavior based on improved RNN.Based on the traditional RNN prediction mechanism for serialized data,an analysis network is designed to predict the serialized data of driving behavior characterization parameters.The network returns the information of the output layer to the hidden layer and carries out bidirectional fusion processing,the test result depends on the information characteristics of the input and output directions,which improves the accuracy of the prediction of the serialized data features.The best parameter selection experiment and the comparison experiment of the prediction effect are carried out for the bidirectional improved RNN,the prediction accuracy of two-way improved RNN is better than that of traditional RNN.(3)Prediction of driving behavior in deep car-following network.The deep car-following network target detection algorithm,the two-way improved RNN serialized data prediction algorithm and the velocity ranging algorithm are fused under Keras to construct the deep car-following network driving behavior analysis algorithm,to predict the driving behavior under the excitation of different driving states of the lead vehicle,an experimental environment was set up to evaluate the prediction effect of the driving behavior under the excitation of the deep car-following network.The experimental results show that the driving behavior of the lead vehicle under the excitation of different driving states can be predicted by the deep car-following network,which can be used to predict driver's driving behavior.