Bayesian inference of multiple object classifications through disparate classifier fusion

This work examines the problem of multiple object classification using disparate sensors where the correct independent classification of all objects is either impossible or requires significantly more measurements than fusing measurements on different objects. It is assumed that the total number of objects being classified is known, but the number of objects in each class is not known. An empirical Bayesian method is employed to first estimate the number of objects in each class using measurements from disparate classifiers, and then fuse these estimates into an estimate over all object classes via Dempster's rule of combination. Using this estimate, a second inference proceeds over the categorically distributed classification probability mass functions for each object. The estimated number of objects in each class is used as a model parameter during this inference. It is shown that by fusing classifier outputs, the classification of multiple objects converges significantly faster to the correct classifications than when inference proceeds independently on each object.