Abstract

Adolescents in resource-limited schools are increasingly affected by ergonomic problems that arise from prolonged computer use and inadequate classroom equipment. Existing posture monitoring systems often rely on expensive sensors or large, fully labeled datasets, which makes them difficult to apply in ordinary educational settings. To address this challenge, the present study introduces a lightweight semi-supervised framework that identifies sitting postures using only a standard webcam. The approach combines expert guided clustering with iterative pseudo labeling so that a very small set of manually annotated images can be expanded into a more comprehensive training dataset. Starting from ten labeled samples, four augmented datasets were created by adding pseudo labeled images generated through different clustering strategies. Four machine learning models, including XGBoost, LightGBM, Random Forest, and a simple artificial neural network, were trained and evaluated using accuracy, F1 score, AUC, and training time. LightGBM provided the most practical balance between predictive performance and computational efficiency, achieving an F1 score of 0.9805 with the shortest training duration. Although Random Forest achieved the highest F1 score, its long training time limited its suitability for real time use. The proposed framework demonstrated stable generalization when tested on students who were not part of the training set, indicating robustness across individual posture differences and varied camera positions. The results suggest that this method can serve as an affordable and scalable tool for everyday posture monitoring in real classrooms, supporting early detection of unhealthy sitting patterns and promoting safer digital learning environments.