Multi-Modal Human Identification System

In this paper we describe our system for multi-source human identification. The system includes a collection of classifiers that classify feature streams derived from audio and video sources. We combine outputs of individual classifiers within our Approximate Bayesian combination framework. The system gives its prediction of the user identity in stantaneously, whenever any useable measurement becomes available. That leads to almost 100% video frame utilization. That is, some prediction is available for almost every frame of the test video. The system is distributed across several machines running independent feature classifiers on the subscription basis. This architecture allows us to successfully use a heterogeneous network of computers regardless of their architecture and operating system. The system has undergone testing in an office environment and shows promising results with respect to increased accuracy and robustness of the classifier combination.