Reaction-Time-Based Driver Sleepiness Estimation via Robust Polynomial Regression for ADAS

Driver drowsiness impairs vigilance and increases crash risk, making real-time monitoring vital for Advanced Driver Assistance Systems (ADAS). While Detection Response Task (DRT) reaction times correlate with Karolinska Sleepiness Scale (KSS) scores, existing mapping techniques often suffer from overfitting and poor generalization due to unregularized high-degree polynomials and ad hoc filtering. This study presents a regression framework for estimating KSS from DRT data, integrated into a multimodal fusion and embedded-systems architecture. The pipeline applies quantile-based scaling, RANSAC-based outlier removal, low-degree polynomial expansion, and Ridge regularization, with hyperparameter tuning via GridSearchCV. Evaluation on data from 20 drivers across 80 simulated sessions showed a 17 % reduction in mean absolute error and a 20 % drop in mean squared error, enhancing prediction accuracy. Correlation coefficients between predicted and actual KSS more than doubled, confirming stronger alignment and increased reliability for real-time monitoring. Multimodal fusion further improved detection sensitivity and reduced false alarms by <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{2 5 \%}$</tex>. Distinct model coefficients across users indicate potential for personalized calibration. Overall, the results validate the feasibility of lightweight, interpretable, and scalable sleepiness estimation models for embedded ADAS, enabling continuous, real-time driver monitoring in safety-critical environments.