Designing an advanced early warning system to detect drivers’ distraction and drowsiness – A pilot study

Abstract Driver distraction and drowsiness are leading causes of traffic accidents. However, current monitoring systems often rely on fixed, population-level thresholds that limit generalizability and increase false-positive rates. This pilot study develops and validates a non-invasive, camera-based driver monitoring system employing multimodal detection of distraction and drowsiness through concurrent assessment of five cephalic and facial area (CFA) parameters: right and left pupil position, head tilt, and right and left eye closure. In this pilot study twenty healthy adult participants completed three driving simulator scenarios — reference (concentrated driving), distraction (deliberate diversion of attention), and drowsiness (simulated fatigue) — while multi-perspective near-infrared imaging continuously recorded CFA parameters. We employed Gaussian probability density function fitting with sigma-based confidence intervals to establish individualized, participant-specific detection thresholds rather than population-averaged norms, enabling simultaneous, independent assessment of distraction and drowsiness states through a decoupled detection architecture. The individualized sigma-based calibration yielded a robust 6–8-fold sigma differential between reference states (sigma ≈ 0.19–0.24) and disstracted states (sigma ≈ 1.54–1.85), substantially exceeding typical fixed-threshold approaches. Detection latency was optimized to less than 1 s for distraction and less than 500 ms for drowsiness, aligning with the critical 5–30 s intervention window identified in naturalistic pre-crash video analysis. The system successfully detected concurrent distraction and drowsiness events through decoupled parallel processing, addressing a documented limitation in sequential or single-classifier systems. This pilot study demonstrates that individualized sigma-based threshold calibration combined with concurrent multimodal detection provides a methodologically rigorous framework for non-invasive driver state monitoring, substantially reducing false-positive rates attributable to inter-individual anatomical variation while maintaining detection accuracy comparable to or exceeding contemporary single-modality systems. Future research will validate these findings in real-world driving conditions with larger, more diverse participant samples.