Application of risky driving behavior in crash detection and analysis

Abstract

Traffic crash detection is a promising and challenging research topic. Due to the limitations of data collection, previous studies mainly used traffic flow variables to establish a traffic crash detection model, and the contribution of risky driving behavior to the traffic crash detection model was not clear. The widespread application of traffic detectors and in-vehicle AutoNavigator software make it possible to collect and update real-time traffic flow data and risky driving behavior data in a short period of time. These data lay the foundation for this study, which aims to quantify the improvement degree of risky driving behavior in a traffic crash detection model and then analyze the coupling effect of risky driving behavior and traffic operation state on the impact of traffic crashes. In this research, we investigated real-time and dynamic traffic flow data and risky driving behavior data by using eXtreme Gradient Boosting (XGBoost) and the logistic regression algorithm, respectively. In addition, SHapley Additive exPlanation (SHAP) was employed to analyze the results and the importance of individual features. The results indicate that the model with the combined inputs has increased accuracy of 8% and nearly a 5% reduction in the false alarm rate. The results of feature importance analysis show that in the variables of risky driving behavior and traffic flow, the most important feature influencing traffic crashes is sharp deceleration. In addition, the characteristics of risky driving behavior increase or decrease the probability of traffic crashes caused by traffic flow characteristics. The results of this paper will help with real-time crash detection and relevant policy-making.