Gait trajectory prediction using Gaussian process ensembles

The development of robotic devices for the rehabilitation of gait is a growing area of interest in the engineering rehabilitation community. The problem with modelling gait dynamics is that everybody walks differently. The approach advocated in this paper addresses this issue by modelling the gait dynamics of individual patients. Specifically, we present a model learner which performs automated system identification of patient gait. The model learner consists of an ensemble of multiple-input-single-output Gaussian Processes which feature automatic relevance determination kernels for automated tuning of parameters. First, the paper presents results for the application of the Gaussian Process ensemble to the learning of a particular patient's gait using a typical prediction configuration. Generalisation of gait prediction is tested with multiple patients and cross-validation. Finally, initial results are presented in which the Gaussian Process ensemble is shown to be capable of learning the mapping between the patient's gait and the therapist-assisted gait.