Learning Tactic-Based Motion Models of a Moving Object with Particle Filtering

Learning motion models of a moving object is a challenge for autonomous robots. We address the particular instance of parameter learning when tracking object motions in a switching multi-model system. We present a general algorithm of joint parameter-state estimation based on multi-model particle filter. We apply the approach to a specific ball-tracking problem and extend the algorithm to learn model parameters in a dynamic Bayesian network (DBN). We show empirical results in simulation and in a team robot soccer environment, as a substrate for applying the learned models to object tracking in a team. The learning capability allow the tracker to much more effectively track mobile objects.