Multi-layer joint gait-pose manifold for human motion modeling

We present a multi-layer joint gait-pose manifold (multi-layer JGPM) for human motion modeling to enhance the representative capability of the original JGPM that represents gait kinematics by two variables. One is the pose to denote a series of stages in a walking cycle and the other is the gait to reflect the individual walking styles. Coupling pose and gait variables in the same latent space was shown effective for human motion estimation. However, the original JGPM is limited to one kind of human gaits, and its learning cannot be scaled up to a large dataset due to a high computational load. This work overcomes the limitations of the previous method by involving a multi-layer topology prior that is able to accommodate a variety of walking styles, leading to better motion synthesis results. Moreover, to learn multi-layer JGPM effectively and efficiently, we adopted two techniques, training data diversification and topology-aware local learning. The experimental results confirm the advantages and superiority of our proposed method over several existing Gaussian process-based motion models.