Analysis On Driving Behavior And Prediction On Driving Style Based On Subjective And Objective Data

Traffic accidents have caused a lot of casualties and economic losses.More than one million drivers are killed and 50 million injured in traffic accidents every year,and the economic losses caused by traffic accidents amount to trillions of yuan.Most of traffic accidents could be attributed to high-risk driving behaviors of drivers,such as violations,mistakes,and distractions,etc.Drivers who often take high-risk driving behaviors have more aggressive driving style,and aggressive drivers are more likely to cause traffic accidents.It is helpful for reducing traffic accidents to train aggressive drivers before driving or intervene in aggressive driving behaviors while driving,thus reducing casualties and economic losses,and then improving overall traffic safety.Therefore,based on subjective and objective data,researches on driving behavior analysis and driving style prediction have been carried out in the paper.The main research contents are as follows:(1)The influence of drivers’ subjective factors on high-risk driving behaviors has been analyzed.Based on the large-scale subjective questionnaire of the Highway Strategy Research Project II(SHRP 2),a Structural Equation Model(SEM)has been constructed to explore the relationship between driving experience,psychological factors(sensation seeking,risk perception)and high-risk driving behavior.The results show that driving experience has significant influence on psychological factors,meanwhile psychological factors have significant influence on high-risk driving behaviors;And then,a mediation model has been constructed based on the SEM to explore the mediating effect by psychological factors.The results show that driving experience which completely mediated by psychological factors,indirectly affect high-risk driving behavior;In addition,a moderating model has been constructed based on the mediation model to explore the moderating effect of gender characteristics.The results show that the driver’s gender characteristics moderates the path from driving experience to high-risk driving behavior: Compared with male drivers,the accumulation of driving experience of female drivers inhibited the high-risk driving behavior more obviously.(2)The prediction model of driver’s driving style has been constructed.The drivers’ crash and near-crash(CNC-rate)of SHRP 2 natural driving data which clustered by K-means algorithm,are used as the driving style label of drivers;Meanwhile,drivers’ subjective factors,such as demographic characteristics and psychological factors,and high-risk driving behaviors are used as model inputs,to train and test the Random Forest(RF)model.The results show that the accuracy of the driving style prediction model based on RF algorithm reaches 90%,which is more than 10% higher than other commonly used machine learning algorithms,such as support vector machine(SVM),Naive Bayes(NB)and logistic regression(LR).(3)The influence of objective driving environment factors on lane-changing behavior has been analyzed.18 volunteer drivers are invited to participate in the simulated driving experiment.Vehicle state parameters and driver operation information are collected by the simulated environment of different weather conditions(sunny weather,foggy weather)and road types(highway,urban);1172 lane-changing driving behavior samples are extracted from the original driving segments based on vehicle lateral displacement;The influence of driving environment on lane-changing driving behavior has been investigated by using independent sample Student Test.The results show that there are obvious differences of vehicle lateral state(lateral acceleration,lateral jerk,yaw angle velocity,yaw angle acceleration)among different weather conditions and road types: The lateral acceleration and lateral jerk of highway road and sunny weather are greater than that of urban road and froggy weather.(4)The prediction model of lane-changing driving style fusing the driving environment has been constructed.Based on the vehicle state parameters and driver operation information collected by simulated driving experience,the statistical characteristics of vehicle state parameters of lane-changing behaviors in different driving environments which clustered by Gaussian Mixture Model(GMM),are used as labels of lane-changing driving styles;Driver operation information(throttle,brake pedal position,steering wheel angle)and driving environment characteristics are fused as model input,to train and test the Long Short-term Memory(LSTM)model.The results show that the accuracy of lane-changing driving style prediction model reaches 93%,which is 3% higher than that of LSTM model without fusing driving environment,and 10% higher than that of Convolutional Neural Network(CNN).