A Machine Learning Approach for Next Step Prediction in Walking using On-Body Inertial Measurement Sensors

This thesis presents the development and implementation of a machine learning prediction model for concurrently aggregating interval linear step distance predictions before future foot placement. Specifically, on-body inertial measurement units consisting of accelerometers, gyroscopes, and magnetometers, through integrated development by Xsens, are used for measuring human walking behavior in real-time. The data collection process involves measuring activity from two subject participants who travel an intended course consisting of flat, stair, and sloped walking elements. This work discusses the formulation of the ensemble machine learning prediction algorithm, real-time application design considerations, feature extraction and selection, and experimental testing under which this system performed several different test case conditions. It was found that the system was able to predict the linear step distances for 47.2% of 1060 steps within 7.6cm accuracy, 67.5% of 1060 steps within 15.2cm accuracy, and 75.8% of 1060 steps within 23cm. For separated flat walking, it was found that 93% of the 1060 steps have less than 25% error, and 75% of the 1060 steps have less than 10% error which is an improvement over the commingled data set. Future applications and work to expand upon from this system are discussed for improving the results discovered from this work.