Driving Behavior Classification and Risk Quantification via Multisensor Machine Learning: From Simulation to Reality

Driver behaviours such as aggression and distraction are leading contributors to traffic accidents worldwide. This paper presents a real-time driver behaviour detection system utilizing a multi-modal sensing approach that combines in-vehicle sensors (CAN bus signals, IMU, ADAS perception outputs) with low-cost physiological measurements (heart rate, steering wheel grip force). The proposed framework employs machine learning models to classify and quantify safe, aggressive, and distracted driving behaviours using Gaussian Support Vector Machines (SVM) for classification and Gaussian Process Regression (GPR) for risk-level estimation aligned with accident probability metrics. A comprehensive analysis of feature selection, sensor combinations, and optimal time windows shows that reliable detection can be achieved using 3–12 seconds of driver observation, supporting online deployment. To validate the generalizability of the approach beyond simulation, a custom-built data logger was deployed in real-world controlled driving scenarios that replicated aggressive and distracted behaviours under safe and ethically approved conditions. The models trained solely on simulator data achieved over 92% accuracy in detecting aggressive driving and more than 88% accuracy for distraction detection when applied to real-vehicle data. These results confirm the robustness of the proposed system for practical implementation in driver assistance and telematics applications.