Online stability estimation based on inertial sensor data for human and humanoid fall prevention

Abstract

Distinguishing between dynamically stable and unstable body poses during the execution of whole-body motions is of equal importance for humanoid robots and humans assisted by robotic exoskeletons. In this work, we present a study for developing a real-time system for detecting dynamic instability based on a small number of body-mounted inertial measurement units (IMUs). To this end, we systematically evaluate different online capable classifiers, operating on the data of 1 to 6 body mounted sensors, trained on a dataset of 50 disturbed motions with nearly 30,000 motion frames recorded at 100 Hz. In contrast to the majority of related studies, our system does not make use of thresholding certain sensor values but instead uses machine learning techniques to detect characteristics and patterns of features of unstable movements. We show that the right combination of classification method and sensor placement on the human body leads to very good detection results with only 3 sensors.