Multi-Source Human Identification

In this thesis, a multi-source system for human identification is developed. The system uses three sources: face classifier, height classifier, and color classifier. In the process of developing this system, classifier combination and the integration of classifer outputs over sequences of data points were studied in detail. The method of classifier combination used relies on weighing classifiers based on the Maximum Likelihood estimation of class probabilities. The integration of classifer outputs, which is termed “temporal integration” in this thesis, has been developed to take advantage of the information implicitly contained in data correlated through time. In all experiments performed, temporal integration has improved classification, up to 40% in some cases. Meanwhile, the method of temporally integrating the outputs of multiple classifiers fused using our classifier weighting method outperforms all individual classifiers in the system. Thesis Supervisor: Tomaso Poggio Title: Eugene McDermott Professor Brain Sciences Department McGovern Institute and Artificial Intelligence Lab