SBIR/STTR Award attributes
Hypersonic technologies are rapidly emerging into many defense platforms and are compounded by a congruent emergence of smart autonomy. In parallel, these technologies will allow for sophisticated high-speed maneuverable vehicles and missiles. As the United States maintains strategic advantages on these technologies, it is imperative that the U.S. also maintain superior counter technologies, including the ability to detect, track, classify, and extrapolate hypersonic targets of interest for early warning systems. Naturally, existing overhead persistent infrared (OPIR) sensors provide an excellent medium to apply state-of-the-art machine learning and/or artificial intelligence technologies for hypersonics but there are a numerous challenges. First, conventional machine learning techniques rely on detailed image features for classification whereas OPIR sensors are far from earth and targets will appear small (e.g. 2 pixels) in images, thus lacking clearly defined static image features. This will require the adaptation of current machine learning techniques to pull classifications features from time dependent target kinematics and behaviors rather than relying on just a single image. Second, environmental background clutter conditions in the OPIR will vary wildly due to differences in the large terrain covered by the sensor as well as effects of clouds, glint, sensor noise, etc., yet performance characteristics must achieve a minimum baseline of consistency. Most importantly, however, is that all empirically trained parametric algorithms like neural networks require a representative training dataset to achieve performance. To address the emerging needs of hypersonic target and track in low earth orbital (LEO) OPIR systems, EpiSci’s algorithm team has partnered with Raytheon Missile Systems’ modeling and simulation team to propose AIded HyperTRACE, a modular tactical AI detection, classification, and predictive tracking system. The approach aims to use Raytheon’s 15 years and more than $10M in high fidelity IR and sensor simulation development to produce representative training and evaluation datasets of hypersonic targets in complex environments and adverse conditions. The datasets will be used to train and replace conventional detection, classification and tracking modules with state-of-the-art modular neural network components efficiently and rapidly with multiple evaluation stages throughout program execution.
