H. Kamvoussioras, T. Garefalakis, E. Michelaraki, C. Katrakazas, G. Yannis

Pages: 121-142

Abstract
Predicting risky driving behaviour in real time is crucial for road safety, allowing for early intervention to prevent crashes. Drivers use cognitive processes like attention, perception, and decision-making to navigate changing road conditions. These processes are prone to errors due to cognitive overload or external distractions, leading to risky behaviour. By predicting such behaviour, safety systems can intervene promptly, reducing crash risks and enhancing road safety. The European Horizon2020 project “i-DREAMS” aimed to address these issues by defining, developing, testing, and validating a “Safety Tolerance Zone (STZ)” to ensure drivers operate within safe boundaries. Within the i-DREAMS framework, three levels of driving behaviour were established: “Normal” driving, “Dangerous” driving, and “Avoidable Accident” driving. Many current risk prediction models lack robustness and interpretability, with common black-box approaches providing limited insight into the factors contributing to risky behaviour. This study aims to address these issues by developing a framework that uses large-scale data from intelligent sensor systems and machine learning techniques to create models that are both robust and interpretable. Specifically, four classification models: Ridge Classifier (RC), Support Vector Machines (SVM), Random Forests (RF), and eXtreme Gradient Boosting (XGBoost) – were developed. These models were utilized to categorize driving behaviour into three defined levels of “Safety Tolerance Zone (STZ)”. Also, to improve interpretability, SHAP (SHapley Additive exPlanations) values were employed, providing a way to identify the key features driving individual predictions. The results revealed that the RF and XGBoost models achieved high accuracy, reaching 95% in prediction accuracy. By identifying the factors that influence risky driving behaviour, this framework offers valuable insights for guiding safety interventions, ultimately contributing to enhanced road safety.
Keywords: driving behaviour identification; Ridge Classifier (RC); Support-vector machine (SVM); Random Forest (RF); eXtreme Gradient Boosting (XGBoost)