AttentionAR: AR Adaptation and Warning for Real-World Safety via Attention Modeling and MLLM Reasoning

DOI

Augmented Reality is rapidly transforming everyday experiences, yet its deployment in dynamic, real-world settings often undermines user safety and situational awareness. Existing AR interfaces typically employ static designs that fail to adapt to fluctuating environmental conditions and shifts in user attention. In this paper, we introduce AttentionAR, a proof-of-concept system whose primary contribution is the novel integration of real-time attention monitoring with contextual hazard assessment to dynamically adjust AR overlays. Our approach comprises three complementary modules: an Attention Awareness module that processes physiological and behavioral signals to distinguish between internal and external attention; a Scene Awareness module that leverages multi-frame image analysis and chain-of-thought reasoning via multimodal language models to evaluate risks; and an AR Adaptation and Warning module that modulates interface transparency and delivers timely alerts. Our user study provides initial evidence that this integrated approach can enhance situational awareness and mitigate distraction-related risks compared to static interfaces. We conclude by discussing the implications of our findings, including the challenges of model generalizability that highlight a critical need for personalization, positioning AttentionAR as a foundational step toward more robust, attention-aware safety systems.