Individual Student Attention Detection in Face-to-Face Classrooms Using Multimodal Facial and Wearable Data

Accurately assessing individual student attention in real-world classrooms remains a significant challenge due to the limitations of subjective teacher observation and the inherent complexity of human behaviour. This work presents a robust multimodal framework that integrates synchronised facial video and wearable sensor data to objectively and continuously estimate student attention in face-to-face classroom settings. Our system combines visual, physiological, and motion cues for fine-grained attention prediction by leveraging a deep neural architecture with modality-specific encoders and advanced fusion techniques. Evaluations on the large-scale DIPSEER dataset demonstrate that our approach outperforms single modality baselines and achieves high precision in regression and classification tasks (root mean squared error of 0.87 and one-off accuracy of 76.3%). The results highlighted the effectiveness and generalisability of multimodal fusion for attention detection, providing a scalable and reliable solution for automated classroom analytics.