Modeling And Risk Analysis Of Typical Risky Driving Behaviors On Highway And Urban Expressway

The fatality rate of traffic accidents on highways and urban expressways ranks first in highways and urban roads,and traffic safety is still a challenge.Statistics conducted by National Highway Traffic Safety Administration(NHTSA)show that the total number of traffic crashes caused by or related to drivers’ risky driving(e.g.,errors and violations in the process of perception,decision-making and control)accounts for 74%.How to effectively control such risky driving behaviors has also been a major challenge in the field of road traffic safety for many years.Risky car-following and risky lanechanging behaviors are the two typical types of risky driving behaviors on highway and urban expressway.In-depth exploration of the occurrence and development of risky driving behaviors,analyzing the causes of risky driving behaviors,predicting risky driving behaviors,and exploring the risk evolution of risky driving behaviors are crucial to preventing risky driving behaviors,improving active traffic safety management and control capabilities,and reducing the number of traffic crashes and casualties.Based on the Shanghai naturalistic driving dataset and high D trajectory dataset,this paper firstly extracted car-following events and lane-changing events on highway and urban expressway.Then,a SMo S was proposed to classify the risky driving behavior.Besides,the causality analysis model,which considered the driver’s heterogeneity,was established to analyze the influencing factors(e.g.,traffic environment characteristics,behavioral characteristics)of risky driving behavior.In addition,a real-time risky driving prediction model,which combined deep learning approach with machine learning algorithms,was proposed to predict the risky driving dynamically.Lastly,this paper has investigated the driving risk evolution mechanism of risky car-following,and predicted the risk disturbance of traffic flow caused by risky lane-changing behavior.This paper has investigated key issues such as the classification,influencing factors analysis,real-time prediction and driving risk evolution of risky driving,forming a complete chain of the occurrence and development of risky driving behaviors.The main research contents include the following aspects:Firstly,a total of 16,905 car-following events and 4,083 lane-changing events on highway and urban expressway were extracted from the Shanghai naturalistic driving dataset and the high D trajectory dataset.Considering the mechanism of rear-end crash,vehicle movement trends,and the parameter’s distribution characteristics of driving behavior characteristics and vehicle operating characteristics,a new SMo S,named rearend crash risk index(RCRI),was then proposed to quantify rear-end crash risk using Monte Carlo simulation algorithm.On this basis,the fault tree algorithm was applied to quantify the driving risk of lane-changing behavior.The fuzzy C-means clustering algorithm was finally applied to classify the risky car-following and risky lanechanging maneuvers.The related research lays the foundation for the causality analysis,real-time prediction,and risk evolution research of risky driving behaviors.Secondly,based on the extracted risky car-following and lane-changing events,the characteristic variables of the traffic environment and driver’s behavior were extracted,and the correlation analysis method was used to screen the modeling variables.On this basis,considering the heterogeneity of drivers and behavioral patterns,the case-control conditional Logit model and the random parameter Probit model were employed to analyze the influencing factors of risky car-following and risky lane-changing behaviors.Significant variables leading to risky driving behaviors were found,and the influence of each variable disturbance on the occurrence probability of risky driving behaviors was quantitatively analyzed.Thirdly,the paper has investigated the effects of different observation time windows,different prediction time windows and different prediction models on the prediction of risky driving.The optimal time window length and prediction model combination scheme for real-time risky car-following behavior prediction was obtained in this study.Besides,considering the recognition of driver’s lane-changing intention,a hybrid framework which combined deep learning approaches and mainstream machine learning algorithms was paoposed to predict risky lane-changing behavior.A lane-changing behavior intention recognition model based on Long-short term memory(LSTM)neural network was proposed in this study.In addition,using the machine learning interpretation algorithm SHapley Additive ex Planations(SHAP)and the feature importance coefficient analysis method,the key variables affecting the prediction accuracy of risky car following and risky lane changing behaviors were finally obtained and analyzed.At the same time,this study also analyzed the impact of key variable interactions on risky car-following prediction results.Finally,a trajectory clustering algorithm-spectral clustering-was conducted to classify the driving risk evolution patterns of risky car-following behaviors.The random parameter multinomial Logit model with heterogeneity in means and variances was then applied to analyze the influencing factors of different evolution patterns.In addition,the influence range of surrounding traffic flow caused by risky lane-changing behavior(i.e.,risky cut-in and risky cut-out)was determined according to the changes of its operation state.The evolution process of driving risk caused by risky lanechanging maneuver was analyzed.The e Xtreme Gradient Boosting(XGBoost)model was adopted to predict the impact range and intensity of traffic flow caused by risky lane-changing maneuver.The SHAP algorithm was finally applied to plot the partial dependence figure and analyze the influence of the interaction of each key influencing variable on the prediction results of the XGBoost model.This thesis has completed the relevant research content of “Identification –Influencing factors analysis – Real-time prediction – Risk evolution” of risky driving behaviors on highway and urban expressway,which is expected to have significant in driver behavior intervention,development of advanced driver assistance systems(ADAS),and development of road traffic safety operation management and connected vehicle control strategies.