Predicting motion of humans, animals and other objects which move according to internal plans is a challenging problem. Most existing approaches operate in two stages: a) learning typical motion patterns by observing an environment and b) predicting future motion on the basis of the learned patterns. In existing techniques, learning is performed off-line, hence, it is impossible to refine the existing knowledge on the basis of the new observations obtained during the prediction phase. We propose an approach which uses Hidden Markov Models to represent motion patterns. It is different from similar approaches because it is able to learn and predict in a concurrent fashion thanks to a novel approximate learning approach, based on the Growing Neural Gas algorithm, which estimates both HMM parameters and structure. The found structure has the property of being a planar graph, thus enabling exact inference in linear time with respect to the number of states in the model. Our experiments demonstrate that the technique works in real-time, and is able to produce sound long-term predictions of people motion.