Characterization of Hard-to-Acquire Targets using Machine Learning (CHArM)

Due to the proliferation of Low Earth Orbit (LEO) mega-constellations and emergent high-speed threats in the air domain, the United States Space Force (USSF) needs more tracking data to maintain persistent coverage to alert warfighters of threatening behavior. LEO mega-constellations require more tracking data because the spacecraft are typically in the same orbital plane. To enable tactically relevant decision making, the United States Space Force needs the ability to detect, ID, characterize and type hard-to-acquire targets in real-time by capitalizing on the vast amounts of available high-framerate Wide Field of View (WFOV) data. The reduction of high frame rate WFOV imagery to centroids of fast-moving low Signal to Noise Ratio (SNR) objects is impractical on current operational compute infrastructure leveraging conventional techniques. Conventional techniques cannot reduce the frames in real-time. These methods apply thresholds to imagery to find high SNR objects. Computationally complex Multiple Hypothesis Velocity (MHV) techniques can detect low SNR high speed objects but require dedicated supercomputers. These MHV techniques can optimized to drastically reduce the computational complexity and significantly reduce the timelines, but still require expensive compute infrastructure with high power consumption and thermal control. Additionally, these techniques only perform detection and do not provide identification, characterization or typing. These processes are needed to properly disseminate detected objects to the proper processing and exploitation systems, e.g. aircraft vs spacecraft. We propose to develop a computer vision software prototype that can detect, ID, characterize and type hard-to-acquire targets from high frame rate WFOV data in real-time to enable tactically relevant decision making We will develop the software prototype by first leveraging terabytes of existing field test data containing aircraft and spacecraft to create a pristine dataset for training Artificial Intelligence / Machine Learning (AI/ML) techniques. We will develop a software prototype that uses AI/ML techniques to detect, ID, characterize, and type high-speed objects with SNR>6 and SNR<6 in a stack of multiple frames. We will leverage a subset of the developed training data to train our AI/ML models. We will leverage the data not used for training to evaluate the performance of the software prototype. We will regularly demonstrate our progress against these objectives. We will work with AFRL and potential transition partners to develop and document a software maturation and transition plan to deliver this capability to the warfighter.