Real-time Trust Prediction in Conditionally Automated Driving Using Physiological Measures
DOI: 10.48550/arxiv.2212.00607
URL: https://arxiv.org/abs/2212.00607
archive: archived pipeline: cataloged verified
Summary
Driving-simulator study (Ayoub, Avetisian, Yang, Zhou; IEEE T-ITS) of real-time driver trust prediction in SAE Level-3 conditionally automated vehicles under engineered system malfunctions. Participants experienced eight takeover requests across three conditions (control, false-alarm, miss) while galvanic skin response, heart-rate indices, and eye-tracking metrics were recorded. Five ML models were compared; eXtreme Gradient Boosting (XGBoost) achieved the best performance, predicting trust dynamics with an f1-score of 89.1%, supporting in-vehicle monitoring systems that calibrate trust to facilitate driver-AV interaction.
Key finding
XGBoost on combined GSR, heart-rate, and eye-tracking features predicts real-time driver trust in conditionally automated vehicles at f1=89.1%, providing a feasible signal for in-vehicle trust calibration during takeover events.
Methodology
Exp 1: 10 participants, repeated measures across 6 sessions from 26 total. Exp 2: 20 participants, Old/New sequence comparison. On-road driving paradigm with DRT and NASA-TLX measures.
Sample size: Exp 1: N=10; Exp 2: N=20
Quality score: 5 / 5