Evaluating student attention in online lectures is challenging, particularly when cameras and microphones are turned off. In this work, we propose a webcam-only behavioral monitoring system that extracts a 12-dimensional feature descriptor per frame using MediaPipe FaceMesh and Pose landmarks. These temporal features are processed by a two-layer stacked LSTM model to classify attention into three states: Focused, Wandering, and Disengaged. The model is evaluated on the SVFM-850 dataset and achieves an accuracy of 91.6% with strong performance across all classes. Under an identical input representation and participant-stratified split, the recurrent model improves on a frame-level 1-D CNN baseline by 3.8 percentage points, isolating the contribution of temporal context. Results indicate that ocular and cephalic features contribute most significantly to prediction, while labial and postural signals provide complementary support. The system operates efficiently on a standard laptop, making it suitable for real-time deployment in virtual classrooms.

online learning, student engagement, LSTM, behavioral analysis, eye tracking, head pose estimation, webcam-based monitoring