Bayesian Nonparametric Modeling and Transfer Learning for Tracking under Measurement Noise Uncertainty

We propose a method for object tracking under unknown and time-varying environmental conditions that incorporates transfer learning with Bayesian filtering and Bayesian nonparametric modeling. The main tracking task assumes that the sensor measurement noise characteristics are unknown and change with time. The characteristics are learned by incorporating knowledge that was previously acquired and stored by multiple sources tracking under similar varying conditions. We assume that each learning source models their own time-varying noise distribution using Dirichlet process mixtures whose parameters are learned using Bayesian nonparametric priors. The multiple models are transferred and combined to model variation in the main tracking task. Simulations demonstrate the improved tracking performance when compared to tracking without transferred knowledge.