Gaze Betray You: Abnormal Attention Detection from Human-Robot Interaction Dynamics

Abstract

Monitoring abnormal attention has a wide range of applications in Human-Robot Interaction systems, such as autonomous driving, virtual reality, and remote operation. Due to factors like fatigue, drowsiness, and distraction, human attention can deviate from tasks that require focus, resulting in abnormal attention during interactions. Monitoring for abnormal attention is essential to ensure the safety of human-robot interactions. However, existing methods often rely on un-explainable algorithms such as neural networks, making them difficult to trust in scenarios like autonomous driving, where explicit decision-making processes and mechanisms are required. This paper presents EyeAAD, an abnormal attention detection method based on dynamic system theory. This method uses gaze signals from a serious game designed to simulate human-robot interaction. We developed several attention-related features for interpretable modeling using a deterministic learning method. By analyzing the matching relationship between the gaze signal and the predicted status from the learned dynamic models, we can detect abnormal human attention patterns. To validate our approach, we designed a serious game for eye movement data acquisition in human-robot interaction. Our analysis showed that EyeAAD enables efficient, real-time monitoring of human attention.