Steven Szklany, Anduril; Anthony Trezza, Anduril; David Smart, Anduril
https://doi.org/10.64861/QOKZ8157
Keywords: Data association, sensor fusion, angle tracklet generation, space domain awareness, machine learning, visual magnitude, low-cost sensors
Abstract:
Optical tracking of objects in low Earth orbit (LEO) typically relies on the angular resolvability between targets as the key observable for association disambiguation. However, when multiple closely-spaced satellites share a consistent and prolonged angular trajectory, kinematic disambiguation alone can be challenging. The recent proliferation of homogenous LEO spacecraft constellations not only affects the density of objects being observed with similar angular trajectory characteristics, but also reduces the effectiveness of visual magnitude as a disambiguating feature.
This paper develops a photometric likelihood and information-theoretic diagnostic that quantifies when photometric fusion adds information beyond angle-only measurements. A single-hypothesis tracker is adapted to exploit this likelihood function, resulting in a real-time capable, robust angle-tracklet generator. This is achieved by carrying the visual magnitude in a separate, per-tracklet photometric filter that is coupled to the kinematic state through an association decision.
Performance of the proposed single-hypothesis approach is compared against a Multi-hypothesis Tracker (MHT) with and without the photometric likelihood. An MHT was chosen as it is the traditional approach for handling these prolonged ambiguities by delaying the association decision until the ambiguities can be resolved. This results in a tracker that can handle more complex association scenarios at the cost of additional computational complexity required to compute the numerous plausible association decisions. Effectiveness of each approach was evaluated using a V-measure score which encompasses the harmonic mean of homogeneity and completeness of the reported tracks.
A realistic simulated scenario of closely-spaced targets was used to demonstrate the benefits of the proposed solution. Utilizing the photometric likelihood function in the single-hypothesis tracker resulted in a V-measure of 0.803. This is a substantial improvement over the angle-only association tracker, which produced a V-measure of 0.084. Further, the proposed solution recovered all simulated targets as distinct output tracklets at every tested angular spacing. Applying the same photometric likelihood function in an MHT resulted in a V-measure of 0.57, compared to 0.11 for the angle-only MHT. As expected, the measured runtime of the MHT was more than 250 times that of the single-hypothesis tracking approach. This makes the single-hypothesis approach an appealing option for real-time applications that must produce tactically relevant orbital information using constrained computational resources.
Date of Conference: September 15-18, 2026
Track: Astrodynamics