The importance of models in Bayesian data fusion

One source of errors in automatic data fusion systems is examined for the simplest case in which the separate sensors supply independent information. Despite the apparent simplicity of this scenario, improvements in performance can still be made over the currently used methods. A theoretical technique is worked through and an approximation to it assessed. Experimental results are given both for synthetic Gaussian data and for a real data fusion problem involving ship silhouette recognition.<<ETX>>