Efficient Vision-Based Framework for Real-Time Classroom Engagement Assessment

Measuring student engagement is critical for improving learning outcomes, yet traditional methods such as manual observation and wearable sensors are often subjective, intrusive, and difficult to scale. This paper presents a vision-based classroom system that evaluates engagement through face detection, centroid-tracking motion analysis, and deferred identity recognition. Unlike prior approaches that rely on continuous re-identification or dense landmark extraction, our method employs periodic identity assignment and Exponential Moving Average smoothing, achieving significant computational savings without compromising accuracy. Experiments on classroom video footage at Swinburne Vietnam demonstrate a 92% correlation with teacher annotations and real-time operation at 25 FPS, validating both accuracy and scalability. The proposed system provides practical insights into student engagement in higher education, offering a privacy-preserving and computationally efficient alternative to existing observation- and sensor-based approaches.