A data fusion approach for classification from disparate knowledge

This paper is concerned with a classification problem in a multisource context dealing with disparate and imperfect data (imprecise, uncertain, incomplete...). Given an exhaustive set of hypotheses, the question is to recognize the most likely one in a given situation described by a set of feature observations generated by sensors or delivered by data processing. Our approach deals with disparate information which can be modelled using different formalisms in order to take into account their specificities. The fusion process involves the choice of a unified formalism able to handle the available knowledge useful for discrimination purpose. The proposed solution is defined in a federative framework provided by the evidential theory. Within this theoretical framework, decision criteria are derived in order to perform the discrimination task