An Inertial Data-Based Upper Body Posture Recognition Tool: A Machine Learning Study Approach

This study presents the development of an automatic and user-independent human motion recognition tool, using kinematic data from inertial measurement units. Data from 50 heterogeneous individuals were collected, and a benchmark of 4 normalization techniques, 4 dimensionality reduction algorithms and 10 machine-learning classifiers was conducted. The tool with the best performance was achieved using the z-score normalization, the mMRM as dimensionality reduction technique, and a Quadratic SVM classifier. This tool presented an overall accuracy of 94,76% in the recognition of 6 static and 10 transitional postures. This work is relevant for awkward postures identification and can be integrated with established frameworks such as RULA and LUBA for ergonomic risk assessment and workspace redesign.