Approximate Learning and Inference for Tracking with Non-overlapping Cameras

Tracking with multiple cameras requires partitioning of ob servations from various sensors into trajectories. In this paper we assume that the observations are generated by a hidden, stochastic 'partition' process and propose a hidden Markov model (HMM) as a generative model for the data. The state space for the hidden variable is intractable, so the inference and learning in our HMM are based on ap proximate representation of the distribution on this state space. The proposed approximation truncates the distri bution from unlikely states. We test our method on real observations; by tracking people in a university building. The tests show that the described approach is an useful al ternative to the existing approximate methods.