Research On Driving Risk Quantification And Crash Identification For Intersection Cooperative Vehicle-infrastructure Warning

As a convergence zone of traffic flow,intersection is an typical accident prone area due to complex traffic environment.Active safety technologies such as intersection vehicle warning based on Cooperative Vehicle-Infrastructure System can identify potential risks based on information fusion,assist drivers to avoid or mitigate accidents,and improve road traffic safety effectively.However,the Intersection warning technology still has certain limitations in the application of differentiated graded warning,risk analysis and accident identification.In view of this,typical elements are extracted from real cases in CIDAS(China In-Depth Accident Study)database to build a static traffic environment using Pre Scan software,and the dynamic control function in the Cooperative Vehicle-Infrastructure environment is realized combining with Simulink software in this study.Questionnaire survey and driving simulation test are used to obtain the original test data.Secondly,the driver’s style is determined based on spectral clustering and EMW(the Entropy Weight Method)by using the key features of the pre-test,and compared with Kmeans clustering results.The correlation between driver style and collision events is explored,and the differences of drivers’danger prevention behavior are analyzed based on the feature extraction of formal test.Thirdly,a differentiated graded warning strategy for driver styles is proposed based on the risk response time of driver categories under different warning modes.The driving risk level is defined by the index TTC(Time to Collision),the driving risk level prediction model is built,and the interpretable SHAP method is applied to analyze the marginal effect of multiple factors qualitatively.Finally,in order to establish the relationship between the surrogate safety index and the accident,the block maxima extreme value theory is introduced,and the Probability Density Functions(PDF)of generalized extreme value distribution under collision and non-collision conditions are fitted respectively by using the absolute value of deceleration change rate as a surrogate index of safey.The Cumulative Density Function(CDF)comparison diagram and Probability Plot are used to verify the fitting effect of the model.A non-collision and collision event classification model is built,and the sensitivity and specificity indexes are used to determine the optimal threshold,thus realizing the crash identification based on surrogate safety indexes.The correlation between the warning strategy and the crash identification results is further explored.The research results show that:(1)The two categories of drivers divided by spectral clustering and EWM have more reasonable differences in related indicators compared with the clustering result by Kmeans.Compared with category I drivers,category II drivers have lower accident risk and better danger prevention ability.They respond quickly,drive at a lower speed before the intersection,brake vigorously and smoothly,and steer with small amplitude.The analysis of danger prevention behavior provides a basis for setting the time threshold in differentiated graded warning.(2)Different categories of drivers have obvious differences in three-grade warning time of prompt warning,emergency warning and active braking.Compared with classification models such as Logistic Regression(LR),Support Vector Machine(SVM)and Decision Tree(DT),the AUC value of driving risk level classification model built based on Cat Boost reaches 86.39%,showing good classification performance.Study by SHAP found that the factors of drivers,test scenes and vehicle control show significant differences on different driving risk levels.The study of graded warning strategy and driving risk factors provides a scheme for reducing accident risk and improving road safety.(3)The generalized extreme value distribution model fits well the general distribution of surrogate safey index maxima in non-collision and collision events.The AUC value of non-collision/collision classification model reaches 88.6%,and the optimal identification threshold based on the surrogate safety index(absolute value of deceleration rate)is T=6.5m/s~3,whose identification accuracy based on this threshold reaches 83.3%.The warning strategy which can effectively avoid collision accidents is defined by correlation coefficient method.The research provides a new idea for the study of driver category behavior characteristics,driving risk control,differentiated graded warning and accident prediction and other active safety technologies.This study provides new ideas for driver behavior characteristics research based on different driver categories,driving risk assessment and control and development of active safety technologies,such as differentiated graded warning and accident identification.