Multisensor fusion and model selection using a minimal representation size framework

Abstract

This paper addresses the problem of statistical model selection for model-based multisensor fusion problems. The minimal representation size (MRS) criterion is used as a basis for the selection of a minimal complexity model among a class of stored models, and in addition enables the selection of parameterization, scaling, and data subsampling. This use of an information-based criterion results in a "universal yardstick" for model selection which is easily adapted to new combinations of sensors and parameters. Each sensor is characterized by a constraint equation defined in the measurement space of observed sensor data. The search for the best model structure is conducted using a polynomial time hypothesize and test algorithm that uses constraining data feature sets (CDFS) to instantiate environment models. Analytical formulation of the minimal representation size model selection for tactile-visual fusion with an anthropomorphic robot hand is presented.