A multi-kernel based Gaussian process dynamic model for human motion modeling

In this paper, we focus on the problem of human motion modeling. We adopt the probabilistic modeling approach to over come the over-fitting problem in the parameter training process and propose a multi-kernel based Gaussian process dynamic model. First, we will do the dimensional reduction, and the method is the Gaussian process latent variable model. Different from existing modeling method, we introduce multikernel learning into the dimensional reduction process to capture the complex distribution of high-dimensional data. Second, for modeling the dynamic latent variable, we use a multi-kernel learning. We are not give a strong assumption on form of the nonlinear projection mapping and nonlinear dynamic function, our model automatically learn a suitable nonlinear kernel based on the training samples, and it can fit many kind of times series. We demonstrate the effectiveness of our method on the CMU human motion data set. The Experimental results show that our modeling method achieves promising modeling capability and is capable of predict human motion.