Toward Person Authentication with Point Light Display Using Neural Network Ensembles

Abstract

Automatic authentication of persons has been an important problem of security, and several techniques based on multi-modal biometric processing has been developed over the last decade. To provide a biometric person authentication system based on human movement signals, this paper presents optimal neural network classifiers that are composed of local experts to solve the problem of human gender recognition. Two ensembles of neural network classifiers have been proposed to discriminate the human gender by using the information of moving joints of actors. With a database consisting of 13 males’ and 13 females’ movement, and containing 10 repetitions of knocking, waving and lifting movements both in neutral and angry style each, we have compared the results of ensembles to the conventional classifiers such as single MLP, decision tree, self-organizing map and support vector machine. Furthermore, the sensitivity has been calculated for the comparison with the human performance that has been obtained from the same data. The experimental results show that the ensemble models are much superior to the conventional classifiers and human subjects.