Kyle Merry, Sandia National Laboratories; Matthew Dykstra, Sandia National Laboratories; Brian Hacsi, Sandia National Laboratories
Keywords: Cislunar, Electro-optical, Onboard Processing, Machine Learning, Quantized Neural Networks, Convolutional Neural Networks
Abstract:
Cislunar space presents a unique challenge for space domain awareness. Cislunar objects can inhabit a wide array of orbits, and present a variety of sensor-to-object geometries. To adequately understand these challenges, we evaluated a hypothetical electro-optical sensor’s performance across a range of sensor and object orbits including L1/L2 halos, distant retrograde, and near-rectilinear lunar.
To process these data, we simulated multiple onboard processing chains, with either traditional image processing or neural network-based detection algorithms. We studied the performance of these algorithms over a range of sensor-to-object geometries and sensor pointing configurations. For the neural networks, we also explored the effects of computing in reduced-precision quantized arithmetic, which enables low-SWaP onboard processing.
We present an analysis of the challenges and possibilities for onboard processing of electro-optical data for cislunar space domain awareness.
Date of Conference: September 27-20, 2022
Track: Cislunar SSA