Measuring Trust Degradation in AVs: Opportunities and Challenges Using Body Posture and Deep Learning

Abstract

Trust is a critical factor in the adoption of autonomous vehicles (AVs), yet critical errors (e.g., cybersecurity incidents) to AV operations can severely undermine it. Traditional trust assessments rely on subjective questionnaires, which may lack granularity and real-time applicability. This study proposes a novel approach to trust measurement by developing deep learning models using drivers’ body posture data. A driving simulator experiment with 40 participants was conducted across three drives: baseline, error drive (vehicle failures from simulated cybersecurity threats), and post-error. Participants’ body posture data were captured via video and processed to extract the x-y coordinates of key body points. Subjective trust levels were collected from questionnaires at the end of each drive, and were used to train and evaluate Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) models. The LSTM model outperformed the CNN, achieving up to 96% accuracy by effectively capturing temporal patterns in posture linked to trust changes. Results suggest body posture is a viable, real-time indicator of trust, particularly when modeled with temporally sensitive architectures like LSTM. This approach offers a direction for adaptive AV systems capable of responding to dynamic trust fluctuations.