Research On Driving Anger Recognition Based On Information Fusion

Driving anger,called “road rage”,has gradually become a severe social issue,which means driving anger has severely influenced traffic safety.Anger has a negative impact on perception,identification,decision and volition process while driving,which causes driving performance to degrade finally,leading to a traffic violation or an accident.Therefore,it is needed to predict driving anger at a macro level based on the determination of influencing factors for driving anger;to detect driving anger at a micro level based on the physiological features and driving behavior features under anger.The main research work of this thesis is included as follows:(1)Twenty-one groups of driving simulator experiments were conducted for angry driving which was induced by special stimulation scenarios or driver-to-driver interactive driving through multi-simulators.Moreover,30 groups of on-road experiments were performed for angry driving which was induced by various elicitation events including jaywalking,weaving/cutting in line within limited time.Based on these experiments,the influencing factors of driving anger,driver physiology and behavior signals were collected.The study results indicate that the driving anger induction methods based on driving simulator experiments and on-road experiments proposed in this thesis are effective as the hit rate of anger emotion based the two methods can reach 74.60% and 84.37%,respectively.(2)A selected(predicted)model of driving anger based on Multinomial Logit(MNL)and sensitivity analysis of influencing factors were established.The validation results indicate that the average predicting accuracy of the MNL based model for the four-intensity driving anger states can achieve 77.34%.The sensitivity analysis results demonstrate that there are three decisive influencing factors of driving anger,consisting of driving experience,temperament and uncivilized behaviors from surrounding road users which should be paid more attention in driver training and traffic control.(3)Physiological features were extracted in time domain and frequency domain and some of them including BVP and sample entropy of ECG were selected by analysis of variance(ANOVA)for classifying different driving anger states.Then a method of receiver operating characteristic(ROC)curve analysis was utilized to determine the best discriminant threshold of driving anger with different intensity based on those physiological features.Further,the four driving anger states can be precisely validated by combining those discriminant thresholds.Using the same feature extraction and selection methods,9 driving behavior features consisting of 6 operating behaviors features and 3 vehicle motion features were determined to classify different driving anger states.The 6 operating behaviors features include standard deviation of steering wheel angle(SWA_Std),the mean lower quartile value of steering wheel angle(SWAQ1_Mean)and the mean pedaling speed of braking pedal(PSAP_Mean),while 3 vehicle motion features include standard deviation of acceleration(Acc_Std),standard deviation of yaw rate(YR_Std)and the standard deviation of lane position(LP_Std).(4)To improve real-time performance of recognition model of driving anger,sequential forward floating selection(SFFS)algorithm and least square support vector machine(LSSVM)were employed to select the best features to comprise a subset of optimal features for classifying the different driving anger states.Subsequently,a recognition model of driving anger based on bi-level belief rule base(BRB)was proposed with the fusion of personal and environmental features,physiology and driving behavior features.The proposed model was trained and optimized by real samples from on-road experiments.The validation results indicate that the average recognition accuracy of the proposed model can achieve 84.26%,as well as true positive rata(TPR)of 82.71% and positive predicting accuracy(PPA)of 80.14%,which outperforms the performances of another five widely used model including LSSVM,k-nearest neighbor(KNN),Na?ve Bayes classifier(NBC),back propagation neural networks(BPNN)and C4.5(an advanced decision tree algorithm).Futher,an adaptive driving anger recognition method is proposed when considering driver indivadula differences,which can improve the accuracy and generalization ability of the original recongnition model.The results can provide theoretical foundation for developing detection or intervention systems of driving anger states.