An Automatic Labeling Approach Towards Multi-class Sitting Posture Classification Based on Depth-Sensor Data

This study aims to create a non-contact system for recognizing the sitting postures of office workers, applicable to healthy sitting monitoring. Skeletal point data were obtained via a depth sensor-based Kinect device while subjects performed five different sitting postures. Five angles have been calculated that can differentiate these postures. A fuzzy rule-based automated approach using angle values is proposed to label the data. With this method, two different data sets were created using traditional time-based labeling methods. Angular and geometric features were used to classify the depth values, and 99.6% and 98.9% accuracy were obtained with KNN and Adaboost classifiers. The proposed labeling method outperformed the traditional time-based labeling method according to the classification results. This system offers a high-performance solution for promoting healthy sitting habits in office workers and has applications in health monitoring and robot vision.