Assessment of physical rehabilitation movements through dimensionality reduction and statistical modeling

The article proposes a method for evaluation of the consistency of human movements within the context of physical therapy and rehabilitation. Captured movement data in the form of joint angular displacements in a skeletal human model is considered in this work. The proposed approach employs an autoencoder neural network to project the high-dimensional motion trajectories into a low-dimensional manifold. Afterwards, a Gaussian mixture model is used to derive a parametric probabilistic model of the density of the movements. The resulting probabilistic model is employed for evaluation of the consistency of unseen motion sequences based on the likelihood of the data being drawn from the model. The approach is validated on two physical rehabilitation movements.