Manuscript received June 5, 2025; revised June 30, 2025; accepted September 17, 2025; published December 19, 2025.

Abstract—Prolonged sitting in office environments increases the risk of musculoskeletal disorders. This study develops an automated posture classification system that combines MoveNet Thunder v4 upper-body keypoints with feature engineering and lightweight classifiers: AdaBoost, Multi- Layer Perceptron (MLP), and XGBoost. We curated 8041 multi‑view images from 29 participants in controlled settings and expanded variability with synthetic images. Under standard 5‑fold cross‑validation on combined views, MLP achieved the best overall performance (accuracy 94.59%, precision 96.49%, recall 92.52%, F1-Score 94.46), while XGBoost was close behind (accuracy 94.16%) and attained the highest ROC‑AUC (0.986). To assess generalization, we conducted grouped 5‑fold cross‑subject validation: all pose‑based classifiers performed similarly (accuracy ≈ 0.91), with XGBoost showing a modest edge (accuracy 0.921±0.025; ROC‑AUC 0.973±0.009) and no significant differences among models. A pretrained MobileNetV2 image baseline failed to generalize in this cross‑subject setting (accuracy ≈ 0.51; AUC ≈ 0.52), indicating degenerate predictions and underscoring the data efficiency of keypoint representations. These results support pose‑based methods as accurate and practical for near‑real‑time ergonomic monitoring, with XGBoost a sensible default and MLP a competitive alternative. Future work will expand real‑world data, incorporate temporal modeling from short videos, and explore deeper fine‑tuning and transformer backbones under the same cross‑subject protocol.

Keywords—ergonomic sitting posture, computer vision, machine learning, MoveNet, musculoskeletal disorder

Cite: Theresia A. Pawitra, Farida D. Sitania, Aji E. Burhandenny, Muhammad F. Wijaya, Anindita Septiarini, Hamdani Hamdani, and Vauwez S. E. Fareez, "Detection of Ergonomic Sitting Postures in Office Environments," Journal of Image and Graphics, Vol. 13, No. 6, pp. 686-701, 2025.

Copyright © 2025 by the authors. This is an open access article distributed under the Creative Commons Attribution License (CC-BY-4.0), which permits use, distribution and reproduction in any medium, provided that the article is properly cited, the use is non-commercial and no modifications or adaptations are made.