MiE: A microscopic model for real-time group engagement estimation using gaze and posture

Real-time audience engagement monitoring allows for the data-driven adaptation of content and activities. By making adjustments on-the-fly, brand activations, presenters and planners can sustain attention, ensure material remains relevant, and ultimately enhance the overall attendee experience. The present work introduces a computational model for real-time, non-invasive engagement monitoring to enable the data-driven adaptation of presentations and activities. The microscopic engagement model (MiE) estimates engagement for individuals by jointly analysing gaze direction, arousal, and body position, classifying states as engaged (positive/negative) or not engaged. These individual estimates are aggregated to compute real-time group-level metrics. The model supports longitudinal identity tracking of points of interest and audience, and provides engagement indicators at instantaneous, period, and entire event levels. Furthermore, the Real-World Engagement and Attention Dataset (RWEAD) 1.0 is presented to benchmark engagement and attention. Evaluations demonstrate robust performance, with the system maintaining consistent identity tracking and achieving over 91% engagement accuracy in RWEAD.