Using neural networks to predict running speed from a single pelvis-worn imu

Modern wearable fitness trackers contain on-device inertial measurement units (IMUs) that measure user activity metrics. These trackers, however, typically require access to a GPS-enabled device (such as a smartphone) to measure speed and distance. This presents a problem when GPS signal is sparse or when a runner prefers to run without a phone. Trying to estimate speed from raw IMU data using physics-based models is problematic due to integration drift. Here, we leverage deep learning techniques to construct a neural network model that predicts running speed from IMU-derived metrics alone, negating the need for a GPS-enabled device. We trained a neural network model on data captured by a commercial IMU fitness tracker (Lumo Inc.). The final model predicts running speed on a validation set with a 0.10 m/s mean absolute error.