Acute stress and takeover performance in conditionally automated driving: model development and validation across urban road and highway contexts

Abstract

Autonomous driving at SAE Level 3 (L3) allows drivers to disengage from continuous monitoring within a defined operational design domain (ODD), but requires timely responses to takeover requests (TORs) when system limits or faults arise. Because drivers are often out-of-the-loop (OOTL), their momentary state—particularly acute stress—may critically shape takeover safety. This thesis investigates how acute stress affects takeover behavior in conditionally automated driving contexts and develops an acute stress detection model based on physiological signals, using two complementary studies.
Study 1 employed a driving simulation experiment with 30 participants under three levels of acute stress (no, low, high) in both highway and urban road conditionally automated driving scenarios. Photoplethysmography (PPG) and electrodermal activity (EDA) signals were continuously recorded, and subjective ratings from the STAI-Y1 questionnaire confirmed significant differences in perceived stress, validating the induction procedure. Results showed that high acute stress significantly impaired takeover performance, with scenario-dependent effects. In highway driving, high level of acute stress reduced lane-keeping ability and led to more conservative driving behavior, whereas in urban road driving, it mainly impaired lane-keeping ability without clear effects on other indicators. These findings provide guidance for improving the safety design of autonomous driving systems.
Study 2 trained Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) classifiers on the physiological data from Study 1,using leave-one-subject-out cross-validation (LOSO-CV). All models exceeded 86% accuracy; XGBoost performed best (accuracy = 88.0%, AUC = 0.896). Model interpretability pointed to standard deviation of heart rate, mean EDA, and tonic skin conductance level as the most informative features. Unlike conventional baselining against resting states, this work uses a “no stress” automated-driving condition as the baseline, improving ecological validity.
Collectively, the findings demonstrate that acute stress degrades takeover quality in L3 driving and that acute stress can be detected reliably from unobtrusive physiology. The work provides actionable guidance for driver monitoring systems (DMS) to enable real-time acute stress inference and adaptive handover strategies, advancing safer, more human-centered autonomous vehicle design.

Date of Award	15 Jul 2026
Original language	English
Awarding Institution	University of Nottingham
Supervisor	Xu Sun (Supervisor) & Jiang Wu (Supervisor)