ANYA: A Noteworthy Youth Annotation Dataset for Machine Learning-Based Ergonomic Posture Analysis

Adolescents in limited-resource educational settings are increasingly impacted by ergonomic health problems resulting from extended computer usage and insufficient equipment. This paper presents a hybrid semi-supervised learning framework for classifying sitting postures utilizing solely webcam input, thereby obviating the necessity for wearables or specialized sensors. The framework combines expert-driven clustering with pseudo-labeling to enhance a limited collection of labeled samples into a robust training dataset. Beginning with 10 manually labeled images, four datasets were produced by augmenting with 20 pseudo-labeled instances each, utilizing various clustering methodologies. Four machine learning models—XGBoost, LightGBM, Random Forest, and Artificial Neural Networks (ANN)—were trained and assessed based on accuracy, F1-score, AUC, and training duration. LightGBM attained an advantageous equilibrium of performance and efficiency, yielding an F1-score of 0.9805 and the minimal training duration of 24.27 seconds. Although Random Forest achieved the highest F1-score (0.9856), its prolonged training duration (495.21 seconds) restricts its real-time usability. The framework exhibited significant generalization on previously unobserved student data, signifying resilience to individual postural variability and camera angles. The suggested method facilitates scalable and economical implementation in actual classrooms, providing a pragmatic solution for posture monitoring in digital learning settings.