Biometrics under continuous observations

To improve biometric system performance, we propose that an image stream be used as input and statistical features be extracted from the image stream. In this paper, we explain the concept of statistical feature extraction from an input image stream and the problem that arises when this is done. We also introduce classifiers based on this concept; e.g. the mutual subspace method, the multiple potential function classifier, and the kernel mutual subspace method. Experimental results from a comparison of classifiers demonstrate the effectiveness of statistical feature extraction from an input image stream.