Confidence-driven adaptive time window for real-time driver fatigue detection in Level 2-3 autonomous vehicles: a multi-dataset validation study.

Driver fatigue constitutes a critical safety hazard in Level 2-3 (L2-3) conditionally automated vehicles, where the paradoxical demand for sustained supervisory vigilance despite minimal active engagement accelerates cognitive underload and impairs timely takeover readiness. Existing vision-based driver monitoring systems are constrained by fixed temporal analysis windows and binary classifiers that neither quantify prediction uncertainty nor adapt to the heterogeneity of real-world fatigue dynamics, resulting in elevated false alarm rates and poor cross-domain generalization. This study introduces a confidence-driven adaptive time window (CDATW) framework: a closed-loop neuro-computational pipeline that couples a lightweight MobileNetV3-CBAM-BiLSTM spatial-temporal encoder with a Monte Carlo Dropout uncertainty estimator to produce simultaneous fatigue probability and epistemic confidence outputs at each inference step. The confidence signal governs a window controller that contracts observation periods to 5-10 s under high certainty (confidence >0.85) for rapid warning, and extends them to 20-30 s under low certainty (confidence <0.60) to suppress spurious alarms-instantiating the feedback-driven adaptive sensing principle central to neurorobotic perception. The framework was validated on four heterogeneous public datasets (NTHU-DDD, YawDD, UTA-RLDD, and DROZY) under single-dataset, cross-dataset transfer, and mixed-dataset training protocols. Single-dataset accuracy ranged from 88.6 to 91.8% with AUC of 0.92-0.95, while the adaptive mechanism reduced false alarm rates by 35.2% relative to fixed 15-s baselines. The architecture sustains 38-45 FPS on an NVIDIA Jetson Xavier NX automotive embedded platform, and confidence calibration achieves an Expected Calibration Error of 0.078, with high-confidence predictions (>0.9) attaining 95.6% accuracy. These results demonstrate that uncertainty-aware adaptive temporal reasoning embedded in a deployable neurorobotic architecture constitutes a computationally efficient and practically viable strategy for driver state monitoring in L2-3 autonomous vehicles, with broader implications for closed-loop perception in safety-critical human-machine systems.