Combining Evidence in Multimodal Personal Identity Recognition Systems

We develop a common theoretical framework for combining evidence in multimodal personal identity recognition systems. It is shown that many existing schemes can be considered as special cases of compound classification where all the biometric data available is used jointly to make a decision. A sensitivity analysis of the various schemes to estimation errors is carried out to show that the integration scheme developed under the most restrictive assumptions — the sum rule — and its derivatives are least affected by estimation errors. The combination strategies developed are used as a basis of a multimodal face recognition system and experimentally evaluated on the M2VTS database [1].