Gaussian process for human motion modeling: A comparative study

We evaluate recent Gaussian process (GP)-based manifold learning methods for human motion modeling, including our recently proposed joint gait and pose manifolds (JGPMs). Unlike most GP algorithms that involve either one latent variable or multiple independent variables in separate latent spaces, JGPMs define two variables jointly and explicitly in one latent space to represent a collection of gait data from different individuals. We develop a model validation technique to examine these GP-based algorithms in terms of their capability of motion interpolation, extrapolation, filtering, and recognition. Experimental results on both CMU Mocap and Brown HumanEva datasets show the superiority of JGPMs over existing GP algorithms for human motion modeling.