Inertial measurement unit based human action recognition for soft-robotic exoskeletons

Abstract

Absence from work caused by overloading the musculoskeletal system lowers the life quality of the worker and entails unnecessary costs for both the employer and the health system. Soft-robotic exoskeletons offer a possibility to overcome these problems by increasing the system flexibility, not limiting the supported Degrees of Freedom and being simultaneously an actuator and a joint. Since such exoskeletons can only be designed for using power when supporting the wearer, battery lifetime can be increased by covering only those actions for which support is needed. As regards controls, a major difficulty lies in finding a compromise between saving energy and supporting the wearer. However, an action-depending control can reduce the supported actions to only relevant ones and increase battery lifetime. The system conditions include detection of user actions in real time and distinguishing between actions requiring and not requiring support. We contributed an analysis and modification of human action recognition (HAR) benchmark algorithms from activities of daily living, transferred them onto industrial use cases and made the models compatible with embedded computers for real-time recognition on soft exoskeletons. We identified the most common challenges for inertial measurement unit based HAR and compared the best-performing algorithms using a newly recorded dataset of overhead car assembly for industrial relevance. By introducing orientation estimation, F1-scores could be increased by up to 0.04. With an overall F1-score without a Null class of up to 0.883, we were able to lay the foundation for using HAR for action dependent force support.

Keywords

• machine learning
• human action recognition
• wearables
• soft-robotic exoskeletons
• assembly
• inertial measurement unit.