Augmenting Bayes filters with the Relevance Vector Machine for time-varying context-dependent observation distribution

Bayesian filtering often relies on a reduced system state relating to robot internal variables only. The exogenous variables and their effects on the measurement process are then encompassed within a global observation noise model. Even if Bayes filters proved to be robust to such approximations, special care has to be taken to handle some of these exogenous effects, usually by introducing complex observation distributions or rejection rules. No matter how complex these models are, they often fail in dealing with contextual incidence which can hardly be explicitly encoded. This article shows how contextual information can be introduced within the Bayesian filtering framework by coupling a filter with classification and regression probabilistic models. The classification model provides an efficient context-dependent measurement selection mechanism and is specifically trained with respect to the filter estimation performance. This first component is enhanced by the introduction of context-dependent observation noise provided by the regression model. The performance of this is approach is evaluated and compared with other methods in the context of altitude estimation for a UAV.