Dynamic Real-Time Prediction of Drivers’ Risk Perception Using Deep Learning and Driving Simulator Data

This study presents a dynamic, real-time framework for predicting drivers’ risk perception (RP) using deep learning techniques and driving simulator data. RP, a critical factor influencing driving behavior and decision-making, was modeled through a Long Short-Term Memory (LSTM) neural network, leveraging data collected from the University of Kansas National Advanced Driving Simulator (miniSim). Thirty drivers across three age groups participated in car-following scenarios and risk-inducing events, during which vehicular parameters, mental workload (WL) fluctuations, and driver demographics were recorded. RP was quantified using continuous measures derived from braking and accelerator forces. The LSTM model, designed with six look-back timesteps, demonstrated strong predictive performance, achieving an RMSE of 0.2 and MAPE of 3.28%. Analysis revealed that braking intensity, space headway, traffic congestion, WL fluctuations, and demographic factors significantly influenced RP. Younger drivers exhibited more consistent but less sensitive RP patterns, while older drivers showed greater RP variability, often linked to elevated WL. RP levels increased in congested conditions and decreased during sustained low WL, indicating potential complacency risks. The LSTM model closely captured the temporal dynamics of RP across diverse scenarios, underscoring its potential for integration into Advanced Driver Assistance Systems (ADAS) to enable adaptive, personalized interventions. The findings also highlight the need for demographic-specific training programs and inform policy strategies aimed at enhancing traffic safety. Future work will focus on extending model validation to real-world driving contexts and incorporating additional physiological and environmental parameters to further refine predictive capabilities.