Direct Conversion of Inertial Measurement Unit Data to Joint Angles in the Upper Limb Using Deep Learning: The Influence of Sensor Number and Placement

Abstract

The aim of this study is to develop a deep learning modeling framework for direct conversion of raw upper limb inertial measurement unit (IMU) data into humerothoracic and elbow joint angles, and evaluate the impact of sensor number and placement on prediction accuracy. Thirty five healthy participants performed planar upper limb motion tasks as well as four dynamic activities, including reaching, head touching, ball throwing, and waving. Upper limb kinematics were simultaneously obtained using three IMUs and video motion capture. Tasks were performed when IMUs were both self- and expert-placed. Generative adversarial networks (GANs) were trained to predict upper limb kinematics using raw IM..