Stochastic Systems Group  

Nonparametric Bayesian Methods for Tracking Maneuvering Targets
Emily Fox
SSG, MIT
We consider state estimation problems assuming a dynamic system with an unknown set of correlated inputs. One example is tracking a target that is subject to an unknown set of maneuver modes with transitions at unknown times. The specific graph we examine is comprised of two interacting graphs: a hidden Markov model (HMM) and a linear Gaussian state space model. The outputs of the HMM are the unobserved inputs to the state space model. Unsupervised learning of HMM parameters is a difficult problem itself, but that difficulty is compounded when the cardinality of the set of hidden state values (e.g. number of maneuver modes) is not known a priori. This can be done, however, by placing a hierarchical Dirichlet process prior on the hidden states.
In this talk, we present the model and a learning algorithm as well as background on Dirichlet processes.
Problems with this site should be emailed to jonesb@mit.edu