Collision Risk Perception Models Using Physiological and Eye-Tracking Signals

Accurate risk perception is essential for safe driving, particularly in dynamic and high-risk traffic environments. This study develops machine learning (ML)-based user risk perception models using physiological recording systems to assess driving risks across various scenarios. We evaluate model performance at different granularity levels, ranging from binary classification (e.g., risk versus no-risk), to detailed classifications, resulting in more classes (e.g., incorporating collision subject types, pedestrian types, numbers, vehicle speeds and deceleration rates). Results demonstrate that despite increased model complexity, the proposed risk perception models consistently achieve high performance, with accuracy exceeding 0.92 in most cases except for high-granularity models in vehicle collision scenarios. Models trained for pedestrian-related risks outperformed those for vehicle-related risks, indicating a stronger physiological response to pedestrian hazards. Feature importance analysis reveals that electroencephalogram (EEG) signals (midline channels: Cz, Fz and Pz) play a dominant role, with pupil center shift degree (PCD) and pupil center shift magnitude (PCM) as secondary key contributors. In contrast, features derived from skin temperature and electrodermal activity (EDA) exhibit limited importance due to slower response times. These findings suggest that EEG and pupil features are optimal for real-time risk perception models, with heart activity features serve as complementary factors in enhancing model accuracy and reliability. We also discuss practical applications of these models in driver-vehicle interaction and intelligent transportation systems. By integrating physiological data with environmental perception sensors, these models offer a promising approach to enhancing safety in semi-autonomous driving systems.