Novelty Methods for General Online Object Deep (GOOD) tracking

Novelty is a major issue in intelligence, surveillance, and reconnaissance (ISR) missions since most SoTA algorithms assume a closed set of finite object categories and will not adapt to novel environments. For high-value target (HVT) tracking, the current algorithms assume every object we intend to track is within the finite object categories and will either misclassify or ignore objects within novel categories. Novelty Methods for GOOD will apply novelty detection, characterization, and accommodation algorithms into the GOOD system developed in the previous Phase II. These functions will extend the GOOD system to account for new object types or something novel about the environment. This unique opportunity allows us to prove the domain independence of the approach by applying it to high-value tracking and other domains for potential commercialization. Our novelty approach is theoretically based on Extreme Value Theory (EVT), which models the known distribution based on the extrema of what is known, has minimal assumptions, and is domain-independent. The Novelty Methods for GOOD will add the EVT-based approach to the GOOD system, demonstrate the domain independence of the approach, and provide real-world examples on the effectiveness of this approach. The system will be delivered with unlimited rights to the government.