Research on an Attention Detection Model in Speech Scenarios Based on Computer Vision

Online education and remote meetings make it difficult for speakers to gauge audience attention in real time; without direct eye contact and interactive feedback, communication effectiveness often suffers. Studies show that in face-to-face communication only about 7% of information comes from verbal content, 38% from voice tone, and as much as 55% comes from facial expressions and body language (nonverbal cues). Therefore, a vision-based model that can assess audience attentiveness via visual signals without requiring wearables is highly valuable. Building on prior work, this paper systematically reviews key background and methods, and proposes a temporal multi-modal model fusing facial expression and body posture features. We train and evaluate the model on public datasets such as DAiSEE and on self-recorded TED-style presentation videos. The fusion model achieves 86.3% accuracy and a macro-average F1 score of 0.84 in a four-class attention classification task, significantly outperforming a baseline model. Additionally, extensive experiments validate the model's robustness and applicability. Finally, we discuss potential application scenarios, ethical/privacy issues, and future development trends for attention detection technology in smart education and human–computer interaction.