Related Experiment Video
Updated: Jul 9, 2026
07:15
Tactile Vibrating Toolkit and Driving Simulation Platform for Driving-Related Research
Published on: December 18, 2020
Confidence-driven adaptive time window for real-time driver fatigue detection in Level 2-3 autonomous vehicles: a
1School of Computer and Network Security, Chengdu University of Technology, Chengdu, China.
Frontiers in Neurorobotics
|July 8, 2026
Summary
Driver fatigue in automated vehicles is a safety risk. This study introduces an adaptive system that uses confidence levels to adjust monitoring, reducing false alarms and improving driver state detection.
Area of Science:
- Neuroscience and Artificial Intelligence
- Automotive Safety Engineering
Background:
- Driver fatigue poses a significant safety risk in Level 2-3 (L2-3) conditionally automated vehicles.
- Current vision-based driver monitoring systems struggle with fixed analysis windows and binary classifications, leading to high false alarm rates and poor generalization.
Purpose of the Study:
- To develop a confidence-driven adaptive time window (CDATW) framework for real-time driver fatigue monitoring.
- To improve the accuracy and reliability of driver state detection in L2-3 vehicles by quantifying prediction uncertainty.
Main Methods:
- Implemented a closed-loop neuro-computational pipeline using MobileNetV3-CBAM-BiLSTM and Monte Carlo Dropout for fatigue probability and confidence estimation.
- Developed an adaptive window controller adjusting observation periods based on confidence levels (5-10s for high confidence, 20-30s for low confidence).
- Validated the framework on four diverse public datasets using various training protocols.
Main Results:
- Achieved high single-dataset accuracy (88.6-91.8%) and AUC (0.92-0.95).
- The adaptive mechanism reduced false alarm rates by 35.2% compared to fixed 15-s baselines.
- Demonstrated real-time performance (38-45 FPS) on an embedded automotive platform with good confidence calibration (ECE of 0.078).
Conclusions:
- Uncertainty-aware adaptive temporal reasoning in a neurorobotic architecture is effective for driver state monitoring in L2-3 vehicles.
- The CDATW framework offers a computationally efficient and practically viable solution for safety-critical human-machine systems.
- This approach enhances timely takeover readiness and mitigates risks associated with driver underload and fatigue.