Motion Prediction With Gaussian Process Dynamical Models

Propose a motion prediction technique based on Gaussian process dynamical models, which maps the existing motion to a low-dimensional latent space by nonlinear method and models the dynamics of motion with Markov chain in latent space. After this, a smooth latent trajectory is obtained corresponding to the motion. Then, predict the future latent states follow the tail end of latent trajectory to obtain a new one. Finally, map this new latent trajectory back to observation space to implement motion prediction. The new motion frames by prediction follow the motion rule of existing motion. Experiment shows that our method can automatically synthesize longer motion by prediction with existing motion.