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

Xie, Wantong; Xiao, Peng · 2026 · Crossref

DOI: 10.3389/fnbot.2026.1857548

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the critical safety challenge of driver fatigue in Level 2–3 autonomous vehicles, where the paradoxical demand for sustained supervisory vigilance despite minimal active engagement accelerates cognitive underload. Existing vision-based driver monitoring systems are limited by fixed temporal analysis windows and binary classifiers that fail to quantify prediction uncertainty or adapt to heterogeneous fatigue dynamics. These limitations result in elevated false alarm rates and poor cross-domain generalization. To resolve this, the authors introduce a Confidence-Driven Adaptive Time Window (CDATW) framework, a closed-loop neuro-computational pipeline that dynamically adjusts observation periods based on real-time prediction certainty. The methodology employs a lightweight MobileNetV3-CBAM-BiLSTM architecture deployed on an NVIDIA Jetson Xavier NX automotive embedded platform. The system extracts spatial features from facial video frames using MobileNetV3 enhanced with Convolutional Block Attention Module (CBAM) and models temporal dependencies using a Bidirectional Long Short-Term Memory (BiLSTM) network. Crucially, the framework integrates Monte Carlo Dropout to estimate epistemic uncertainty, producing simultaneous fatigue probability and confidence scores. A window controller uses these confidence signals to modulate the observation window: contracting it to 5–10 seconds when confidence exceeds 0.85 for rapid warnings, and extending it to 20–30 seconds when confidence falls below 0.60 to suppress spurious alarms. The system was validated across four heterogeneous public datasets (NTHU-DDD, YawDD, UTA-RLDD, and DROZY) using single-dataset, cross-dataset transfer, and mixed-dataset training protocols. The results demonstrate that the CDATW framework achieves single-dataset accuracy ranging from 88.6% to 91.8%, with Area Under the Curve (AUC) values between 0.92 and 0.95. The adaptive mechanism reduced false alarm rates by 35.2% relative to fixed 15-second baselines in within-dataset settings. In cross-dataset transfer scenarios, the adaptive window reduced false alarm rates by 25.0%, from 19.6% to 14.7%, despite some degradation in confidence calibration. The architecture sustains inference speeds of 38–45 frames per second, meeting real-time automotive requirements, and achieves an Expected Calibration Error of 0.078, with high-confidence predictions (>0.9) attaining 95.6% accuracy. The significance of this work lies in demonstrating that uncertainty-aware adaptive temporal reasoning is a computationally efficient and practically viable strategy for driver state monitoring in conditionally automated vehicles. By embedding confidence-driven feedback directly into the inference process, the system mimics biological attentional gating mechanisms, improving robustness against distribution shifts and varying lighting conditions. This approach offers a deployable solution for enhancing human-machine collaboration safety in L2–3 autonomous driving, with broader implications for closed-loop perception in other safety-critical neurorobotic systems.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms
• truck driver fatigue
• dms validation
• time on task

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: physiological data
• Methodological Resource: validation psychometrics, tool software