Networked and Embedded Systems Laboratory

Tryear

Trmonth

Trnumber

Title

Abstract

This work investigates decision level fusion by ex- tending the framework of subjective logic to account for hidden observations. Bayes rule might suggest that decision level fusion is simply calculated as the normalized product of the class likelihoods of the various classifiers. However, this product rule suffers from a veto issue. The problem with the classical Bayes formulation is that it does not account for uncertainties inherent in the likelihoods exclaimed by the classifiers. This paper uses subjective logic as a rigorous framework to incorporate uncertainty. First, a class appearance model is introduced that roughly accounts for the disparity between training and testing conditions. Then, the subjective logic framework is expanded to account for the fact that class appearances are not directly observed. Rather, a classifier only returns the likelihood for the class appearance. Finally, the paper uses simulations to com- pare the new subjective logic framework to traditional classifier fusion methods in terms of classification performance and the ability to estimate the parameters of the class appearance model.

Filename

File

Urlpdfpaper

Urlsrcpaper

Urlpdfpresentation

Urlsrcpresentation

Urlavmedia

Urldoi

Urlpublisher

Urlgooglescholar

Urlciteseer

Pubin

Pubvol

Pubnum

Pubnum end

Pubpagefirst

Pubpagelast

Pubpagecount

Pubdate

Pubdate end

Pubplace

Publisher

Ispublic

Islabdocument

Miscattributes

Document category

Main research area

Editing document