Human Arm Posture Estimation using only Pressure Sensors Integrated into Handles - A proof of concept
Sebastian Helmstetter, Lorenz Neumann, Jacob Sembritzki, Andreas Lindenmann, Sven Matthiesen · 2025
In many professions, repetitive and sustained tasks exert considerable physical strain on the human body when performed in awkward postures. Currently, there is a lack for methods monitoring individual's postures during their daily work in order to provide early warning of an ergonomic risk to health. We present how a machine learning-based regression model can estimate shoulder and elbow joint angles of human during drilling tasks, solely using data from pressure sensors embedded on the surface of a power tool handle. For model training and validation, the Xsens motion capture system is utilized for the ground truth measurements of the joint angles. The person-specific models are trained for the seven participants using a Gaussian Process Regression (GPR) algorithm. On average, the models attained a mean absolute error (MAE) of 7.3° in estimating the elbow flexion. Given the movement range of 40° to 110° this corresponds to a relative error of 10.4%. A significant variation in hand positioning on the tool handle is identified during repetitive arm postures, even for the same participants. This variation affects the robustness of the models contributing to the measured error. Despite this challenge, the proof of concept is successfully demonstrated.