Transfer Learning with Bayesian Filtering for Object Tracking under Varying Conditions

We propose an algorithm that integrates Bayesian filtering with transfer learning to track a moving object under unknown time-varying environmental conditions. In order to account for measurement noise intensity variations in the primary source, we use multiple learning sources with labeled measurements. For each source, the measurement likelihood is modeled using Gaussian mixtures whose parameters are learned from conjugate priors. Weighted basis combinations of the multiple learned information are then used to model the measurement likelihood of the primary source; the basis weights are learned using a Dirichlet distribution prior. The improved tracking performance of the proposed algorithm is demonstrated for both low and high noise scenarios.