Driving Behavior Standardization And Characteristic Representation

Driving behavior analysis is a hot topic in the field of intelligent transportation and unmanned driving.Establishing proper driving behavior analysis and presentation model is of great significance to the development of personalized vehicle control system.At the same time,the effective analysis of driving behavior can also provide solutions to problems such as driving style identification and abnormal driving behavior detection.Therefore,how to establish the driving behavior analysis model and take it into practical application becomes the key problem.This project proposes to study the actual driving behavior and map the original driving behavior into the standard vehicle operating condition,the Federal Test Procedre(FTP-72),which can eliminate the disturbance caused by different driving environment and utilize the standardization of driving behavior.The Vehicle Test Data(VTD),which collected in actual driving scenarios with manual marked driving style by motor company,is used as the input data.The average of energy spectrum density is selected as the quantitative indicators of driving style in driving behavior standardization.Experiments show that the proposed method can retain and reproduce the original driving style,and provides a standard reference frame for driving behavior analysis.Furthermore,this paper presents a feature representation model of driving behavior based on phase space reconstruction and pre-trained convolution neural network.The model is based on the results of driving behavior standardization.The accelerator pedal sequences of driving behavior standardization is reconstructed by phase space and the reconstruction curve is fed into pre-trained convolution neural network in the form of picture.The output of convolution layer is the characteristic vector of driving behavior.This project uses real-world normal driving behavior collected in four cities in China and abnormal driving behavior data,which was simulated by existing theory.By employing the t-SNE dimension reduction algorithm and Kmeans clustering analysis algorithm in the designed test experiment,the distribution characteristics and calculation characteristics of different driving behaviors are demonstrated,and the effectiveness of the proposed method is verified.In view of the car following scenario,a personalized car-following driver model is established based on the idea of driving behavior standardization and the direct inverse model in learning control theory.VTD and time headway theory are used to build simulated car-following data in case the real-world stylized car-following data is not available.The personalized car-following driver model is tested by real vehicle following scenario simulation.The test results show that the proposed model can well follow the leading vehicle,with its own driving style retain and reproduced.