Answers are grounded in this paper and cite their sources.
Abstract Driver fatigue is a significant road-safety concern, and camera-based monitoring systems often rely on fixed eye-closure thresholds that fail to account for individual facial differences and transient movements. This study proposes DriveGuard , a personalised, confidence-aware temporal multimodal framework for driver-fatigue estimation and graduated alerting. The framework integrates three feature streams: eye-related indicators (Eye Aspect Ratio, PERCLOS, and blink characteristics), mouth-related indicators (Mouth Aspect Ratio and yawning frequency), and head-pitch features. It employs a 45-second per-driver enrolment process for personalised feature normalisation, followed by temporal encoding across four consecutive 3-second windows using statistical summaries and a multinomial logistic regression classifier. A confidence-aware alert mechanism incorporating exponential moving average smoothing, confidence gating, and hysteresis generates three graduated alert levels: normal, advisory, and warning. The framework was evaluated through a seeded computational simulation involving 48 synthetic drivers and 2,496 sequences, using six-fold driver-disjoint cross-validation. The proposed configuration achieved a per-driver macro-F1 score of 77.5% (95% bootstrap confidence interval: 74.7–80.2%), compared with 58.3% for a conventional fixed EAR/PERCLOS threshold-based baseline. Ablation experiments demonstrated improvements from temporal context (+8.5 percentage points), personalisation (+6.7 points), and the head-pitch feature stream (+6.9 points), while the mouth stream provided no measurable improvement. Furthermore, confidence-aware smoothing reduced alert-level transitions from 10.7 to 3.3 per simulated drive, demonstrating improved alert stability at the cost of increased advisory detection delay. Probability recalibration through sigmoid scaling did not improve calibration. Replication across four additional random seeds and sensitivity analyses supported the consistency of the configuration ranking. These findings demonstrate the potential benefits of combining personalisation, temporal modelling, multimodal feature integration, and confidence-aware alerting within a lightweight driver-fatigue estimation framework. However, as the evaluation exclusively uses synthetic data, the reported performance does not establish real-world accuracy or driving safety. Further validation using ethically approved real-world driving datasets and human participants is necessary.
No takes yet. Share an insight, caveat, or question.
Raj Surase (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: