Austin Ibele, Kung Fu AI; Spencer Romo, Kung Fu AI; Brian Williams, Slingshot Aerospace; Steve Kramer, Kung Fu AI; Tate Noster, Kung Fu AI
Keywords: WFOV, camera array, GEO detection, GEO tracking, SatSim, SatNet, AllSky, space object detection, PANDORA, Convolutional Neural Networks, RetinaNet
Abstract:
Wide field of view (WFOV) optical sensors can provide a tremendous amount of satellite observation data. For example, the PANDORA (Persistent AND Optically Redundant Array) system is a 5×9 array of commercial-off-the-shelf cameras located at the Air Force Maui Optical and Supercomputing site. It captures 20°×120° WFOV images of the night sky at a rate of two frames per minute. Identifying satellites in these images is time consuming and error prone when performed manually. Convolutional neural networks (CNNs) are a natural approach for this problem, but standard object detection CNNs are not well suited for detecting objects that are so small and faint. We make adaptations to RetinaNet object detection neural networks and make use of extensive image augmentations to train models for detecting these small objects. Our CNNs achieve max F1 scores against simulation data of 0.90 and we fine-tune our networks to achieve a max F1 of 0.73 against real PANDORA imagery.
Date of Conference: September 27-20, 2022
Track: Machine Learning for SSA Applications