Automated Entity Classification in Video Using Soft Biometrics

In this SBIR project we will develop an automated capability to quickly recognize and classify subjects in video imagery using soft biometrics. The system will have capability to translate video streams into probabilistic biometric metadata about the observed subjects (i.e. “With 82% confidence, subject is John Smith”). We intend to address both the accuracy and speed issues to produce a simple, intuitive and maintainable/upgradeable system which will allow real time and accurate searches of massive biometric archives. In essence our solutions will work by approximating the entire dataset in a small index which can be kept in main memory. At query time, this index is searched in milliseconds for an approximate answer, which is then confirmed by loading a small fraction of the original data from disk. A crucial observation is that if the index approximation has certain properties (the lower bounding lemma) we can guarantee that the result is the same one we would have gotten if we had done the slower brute force search. This system will allow for translation of descriptions of a person to a soft biometric metadata representation, allowing distributed and (possibly disparate) imagers to collaborate in inferring matches.