Alex Geringer-Sameth, Lawrence Livermore National Laboratory; Nathan Golovich, Lawrence Livermore National Laboratory; Kshitij Bhardwaj, Lawrence Livermore National Laboratory
https://doi.org/10.64861/XARW6286
Keywords: Image processing, astrometry, photometry, track-before-detect, image stacking, image simulation, edge computing, space sensing, GPU, Nvidia Orin, Nvidia Jetson, on-board processing
Abstract:
Distributed space-based optical sensors for Space Domain Awareness (SDA) will require onboard processing to reduce image data and identify moving object candidates without continuous ground contact. We developed a locally self-contained pipeline for short-exposure image stacks that performs detector calibration, source detection, astrometric and photometric calibration, and registration to a common sidereal frame.
The pipeline supports image-based track-before-detect using GPU-accelerated KBMOD and a catalog-based method that clusters low-threshold tracklets in position-velocity phase space. We evaluated the pipeline on the ground using downlinked GEOStare SV2 observations of the near-Earth object 2024 MK. The catalog-based method consistently recovered the asteroid as the largest tracklet cluster.
Processing an 11-image sequence spanning 300 seconds required 43 seconds on a Jetson Xavier NX. Separate benchmarks on a Jetson Orin show that KBMOD runtime is effectively linear in the computational workload, defined as the product of stack depth, image area, and number of trial velocities, with a normalization that depends on GPU power mode. Extrapolation to the velocity range used for the catalog-based 2024 MK detection gives a runtime of approximately 2 × 10⁴ seconds.
For the tested dataset, catalog-based processing kept pace with image acquisition, while brute-force track-before-detect requires a more efficient search strategy.
Date of Conference: September 15-18, 2026
Track: SDA Systems & Instrumentation