Optimal weighting of bimodal biometric information with specific application to audio-visual person identification

摘要

A new method is proposed to estimate the optimal weighting parameter for combining audio (speech) and visual (face) information in person identification, based on estimating probability density functions (pdfs) for classifier scores under Gaussian assumptions. Performance comparisons with real and simulated data indicate that this method has advantages in reducing bias and variance of the estimation relative to other methods tried, so achieving a robust estimator of the optimal weighting parameter. Another contribution is that we propose the bootstrap method to compare performances of different algorithms for estimating the optimal weighting parameter, so providing a strict criterion in comparing algorithms of this kind. Using simulated data, for which the pdf is controlled and known, we show that the advantages of the method hold up when the underlying Gaussian assumption is violated. The main drawback is that we have to choose an adjustable parameter, and it is not clear how this should best be done.