Recursive Bayesian Estimation (RBE) is a widespread solution for visual tracking as well as for applications in other domains where a hidden state is estimated recursively from noisy measurements. Although theoretically sound and unquestionably powerful, from a practical point of view RBE suffers from the assumption of complete a priori knowledge of the transition model, that is typically unknown. The use of a wrong a priori transition model may lead to large estimation errors or even to divergence. We propose to prevent these problems, in case of fully observable systems, learning the transition model on-line via Support Vector Regression. In particular, we have proposed an application of this general framework in the context of linear/Gaussian systems, where we dub it Support Vector Kalman (SVK).