Adaptive Takeover Strategy Based on Driver Fatigue State System for L3 Autonomous Vehicles

In Level 3 (L3) autonomous driving, drivers must remain prepared for takeover requests while supervising vehicle operation. This dual responsibility creates challenges in maintaining effective takeover performance, particularly during driver fatigue. We propose an adaptive takeover intervention framework that adjusts methods based on real-time driver fatigue levels, utilizing a non-invasive fatigue detection algorithm validated on the YawnDD dataset. The system achieved $79.3 \%$ accuracy in eye feature detection and $84.2 \%$ in mouth feature detection. We evaluated our dynamic takeover strategy through a driving simulator experiment with 30 participants. Results showed that in the normal group, the visual-audio strategy reduced reaction times by 0.41 s compared to visual-only and 0.17 s compared to audio-only approaches. The fatigue group demonstrated similar improvements, with the visual-audio strategy showing 0.48 s and 0.13 s reductions compared to visual and audio strategies respectively. The visual-audio approach also produced the most concentrated reaction time distribution, indicating consistent takeover performance. These findings demonstrate that combining real-time fatigue monitoring with adaptive multimodal interventions can enhance $\mathbf{L 3}$ autonomous vehicle safety.