Reconstruction of Angular Kinematics From Wrist-Worn Inertial Sensor Data for Smart Home Healthcare

Abstract

This paper tackles the problem of the estimation of simplified human limb kinematics for home health care. Angular kinematics are widely used for gait analysis, for rehabilitation, and more generally for activity recognition. Residential monitoring requires particular sensor constraints to enable long-term user compliance. The proposed strategy is based on measurements from two low-power accelerometers placed only on the forearm, which makes it an ill-posed problem. The system is considered in a Bayesian framework, with a linear-Gaussian transition model with hard boundaries and a nonlinear-Gaussian observation model. The state vector and the associated covariance are estimated by a post-regularized particle filter (constrained-extended-RPF or C-ERPF), with an importance function whose moments are computed via an extended Kalman filter (EKF) linearization. Several sensor configurations are compared in terms of estimation performance, as well as power consumption and user acceptance. The proposed constrained-EKF (CERPF) is compared to other methods (EKF, constrained-EKF, and ERPF without transition constraints) on the basis of simulations and experimental measurements with motion capture reference. The proposed C-ERPF method coupled with two accelerometers on the wrist provides promising results with 19% error in average on both angles, compared with the motion capture reference, 10% on velocities and 7% on accelerations. This comparison highlights that arm kinematics can be estimated from only two accelerometers on the wrist. Such a system is a crucial step toward enabling machine monitoring of users health and activity on a daily basis.