A machine learning approach for stride speed estimation based on a head-mounted IMU

Walking speed in real-life conditions is typically estimated through wearable inertial sensors mounted on waist, lower limbs, or wrists. Very recently, head-mounted inertial sensors are emerging for gait assessment. The present study explores the feasibility of measuring the stride speed with a head-mounted inertial sensor in both laboratory and real-world settings. The developed algorithm exploits a Temporal Convolutional Network for the detection of the gait events and a Gaussian Process Regression for the stride speed estimation. The experimental evaluation was carried out on healthy young participants during both standardised indoor and real-world walking trials. For indoor trials, errors were smaller than previous studies (0.05 m/s). As expected, errors increased at lower speed regimes due to a reduced signals amplitude. During 2.5-hours real-world evaluation, errors were slightly larger but acceptable (0.1 m/s). Reported results are encouraging and show the feasibility of estimating gait speed with a single head-worn inertial sensor.