Learning Predictive Movement Models From Fabric-Mounted Wearable Sensors

The measurement and analysis of human movement for applications in clinical diagnostics or rehabilitation is often performed in a laboratory setting using static motion capture devices. A growing interest in analyzing movement in everyday environments (such as the home) has prompted the development of “wearable sensors”, with the most current wearable sensors being those embedded into clothing. A major issue however with the use of these fabric-embedded sensors is the undesired effect of fabric motion artefacts corrupting movement signals. In this paper, a nonparametric method is presented for learning body movements, viewing the undesired motion as stochastic perturbations to the sensed motion, and using orthogonal regression techniques to form predictive models of the wearer's motion that eliminate these errors in the learning process. Experiments in this paper show that standard nonparametric learning techniques underperform in this fabric motion context and that improved prediction accuracy can be made by using orthogonal regression techniques. Modelling this motion artefact problem as a stochastic learning problem shows an average 77% decrease in prediction error in a body pose task using fabric-embedded sensors, compared to a kinematic model.