Uncertainty of predicting walking and running motions using inertial measurement unit attached to the wrist

Abstract

Wearable measurement systems with single or multiple inertial measurement units (IMU) are being developed for everyday or outdoor gait analysis. Most existing studies about the prediction of gait from a single IMU select the trunk or lower body as sensor placement. Though the smartwatch is one of the most widely used wearable devices, there is a lack of related studies. The purpose of this study is to quantify the accuracy and uncertainty of predicting gait kinematics from a single IMU worn on the wrist and discuss the feasibility of gait analysis with smartwatches. With a Bayesian neural network trained with variational inference, various cases of sensor configuration and population size used in training are compared. The dataset consists of motion capture and IMU sensor signals of 200 subjects. With inverse kinematics, generalized coordinates of a musculoskeletal model were calculated from the motion capture and then used as the motion kinematics. Data augmentation based on virtual sensor calculation and generative model was tried to examine the effect on prediction accuracy and uncertainty. The RMSE of sagittal plane kinematics predicted from the Bayesian neural network was 5.80 degrees, and the aleatoric and epistemic uncertainty of the model could be quantified. The data augmentation methods could reduce epistemic uncertainty, and improve prediction accuracy if there are subjects at most 40. Through the results of the prediction of gait kinematics based on virtual sensors on various parts of the human body, this study could discuss the suitable position for inertial sensors to predict each generalized coordinate.