Evaluating Machine Learning-Based Classification of Human Locomotor Activities for Exoskeleton Control Using Inertial Measurement Unit and Pressure Insole Data

Tom Wilson, Samuel Wisdish, Josh Osofa, Dominic J. Farris

Classifying human locomotor activities from wearable sensor data is an important high-level component of control schemes for many wearable robotic exoskeletons. In this study, we evaluated three machine learning models for classifying activity type (walking, running, jumping), speed, and surface incline using input data from body-worn inertial measurement units (IMUs) and e-textile insole pressure sensors. The IMUs were positioned on segments of the lower limb and pelvis during lab-based data collection from 16 healthy participants (11 men, 5 women), who walked and ran on a treadmill at a range of preset speeds and inclines. Logistic Regression (LR), Random Forest (RF), and Light Gradient-Boosting Machine (LGBM) models were trained, tuned, and scored on a validation data set (n = 14), and then evaluated on a test set (n = 2). The LGBM model consistently outperformed the other two, predicting activity and speed well, but not incline. Further analysis showed that LGBM performed equally well with data from a limited number of IMUs, and that speed prediction was challenged by inclusion of abnormally fast walking and slow running trials. Gyroscope data was most important to model performance. Overall, LGBM models show promise for implementing locomotor activity prediction from lower-limb-mounted IMU data recorded at different anatomical locations.