ENGAGEMENT TRACKING MODEL FOR ONLINE LEARNING ENVIRONMENT

Abstract

Online learning has expanded rapidly, but maintaining student engagement in virtual classrooms remains a persistent challenge. Most existing engagement detection models are trained on Western or East Asian populations and achieve modest accuracy, particularly for under‑resourced settings. We propose an engagement detection model that classifies students into three categories (very engaged, nominally engaged, not engaged) using four interpretable behavioral indicators: facial emotion recognition (seven emotions), drowsiness detection, phone use detection, and distraction detection. Webcam video was collected from 38 tertiary students in Kaduna State, Nigeria, during live online lectures. Frames were extracted and augmented to produce a balanced dataset of 9,000 images. DenseNet121, pretrained on ImageNet, extracts 1,024-dimensional feature vectors, which are then classified by a lightweight Multi-Layer Perceptron (MLP) with two hidden layers and dropout regularization. On our local test set, the model achieves 82% accuracy and an F1-score of 0.88 for the critical "not engaged" class, outperforming an LSTM baseline (80% accuracy). The MLP classifier runs in under 1 ms per frame, enabling real-time feedback, and its predictions are explainable via the specific behavioral cues that triggered each alert. On the public DAiSEE benchmark, both models fail on minority classes due to severe class imbalance, confirming that data quality and class balance are more critical than architectural sophistication. Limitations include the modest sample size (n=38) and the absence of cross-regional validation. We conclude that locally collected, balanced data can substantially improve engagement detection for underrepresented populations, offering a practical, interpretable, and real-time solution for educators in low-resource settings.