Spatio-Temporal GP Model Learning for Intention-Driven Motions

Most human activities and object motions in the real world are intention-driven. Taking advantage of the intention information (e.g., goals and destinations) can produce better motion models and more accurate trajectory prediction in general. Again, compared with the traditional state space models, Gaussian process (GP) based models have more capability to de-scribe complicated motions. This paper proposes a GP regression based approach to model learning and trajectory prediction for intention-driven motions. At first, the conditional kernels are devised by incorporating the known motion intent, from which it follows that the GP models of intention-driven motions are constructed. Then, the times at which the destination is reached, as key parameters for GP models with conditional kernels, are learned online based on the data stream. Finally, in the context of missile tracking, numerical simulations are provided to show the effectiveness of the proposed GP models and the self-learning ability of their hyper parameters for intention-driven motions.