A Contactless Real-Time System to Classify Multi-Class Sitting Posture Using Depth Sensor-Based Data

Abstract

Musculoskeletal disorders are often linked to poor sitting postures, making the assessment of healthy sitting positions crucial. This study develops a system for contactless recognition of office workers' sitting postures, using various classification methods for health applications. Five sitting postures were defined based on medical literature and standards. Thirty subjects held these postures for 30 seconds, with pose data captured via a Kinect device. To overcome challenges like desks and computers, two datasets with different joint points were created. Pose samples were labeled by calculating angles between body parts like legs, hips, and back. Classification methods included Neural Networks, Support Vector Machine (SVM), K-Nearest Neighbors, Naive Bayes, AdaBoost, Decision Tree, Random Forest, and Ensemble Learning (EL). The highest accuracy was achieved by EL and SVM, at 99.8% and 99.7%, respectively. The first and fifth postures were found to be the most comfortable. This system aims to improve sitting behaviors and is useful for health monitoring and robotic vision.