Multi-hypothesis sensor fusion with an epistemic kernel mixture filter

Tools

Parsonage, Benjamin and Vasile, Massimiliano (2026) Multi-hypothesis sensor fusion with an epistemic kernel mixture filter. In: AAS/AIAA Astrodynamics Specialists Conference 2026, 2026-07-26 - 2026-07-30, Hilton Whistler Resort and Spa.

[thumbnail of Parsonage-Vasile-2026-Multi-hypothesis-sensor-fusion-with-an-epistemic-kernel-mixture-filter]
Preview
 Text. Filename: Parsonage-Vasile-2026-Multi-hypothesis-sensor-fusion-with-an-epistemic-kernel-mixture-filter.pdf
Preprint
License: Creative Commons Attribution-NonCommercial 4.0 logo
Download (1MB)| Preview

Abstract

Resident Space Objects (RSOs) may be observed using several alternative measurements methods, such as Light Curves, radar etc. Each method is subject to various sources of uncertainty, such as atmosphere attenuation/degradation effects, temporal/systematic sensor noise, optical aberrations and surface defects, to name a few. Furthermore, the processing of raw measurements to obtain state estimations can result in several possible solutions given inherent ambiguities in the received data. This paper demonstrates that this multiplicity can be treated as an epistemic uncertainty. For cases in which sequential batches of measurements are available from different types of sensor, we present a robust sequential filter that can handle epistemic uncertainty in the received measurements and the model of the dynamics. This filter can represent an uncertain distribution of arbitrary complexity as a Epistemic Kernel Mixture Model (EKMM) and can propagate a full set of distribution functions by treating each kernel individually. The kernel weights of the EKMM are interval valued, allowing the filter to represent multiple non-Gaussian probability distributions simultaneously. Robust Bayesian process is used to update both the weights and their bounds in order to build a family of posterior distributions from each new measurement. Between observations, kernels are automatically adapted in number and size to optimise the propagation of the prior. From indirect inference through propagation, the filter can also be used to estimate unknown model parameters and update the model distribution accounting for epistemic uncertainty. Results are presented for an attitude estimation problem in which the measurement ambiguities and resulting multiplicity in possible attitude states from several available sensor types is modelled using the proposed filtering approach. It is shown that the continuous propagation of epistemic distributions allows the sequential inference of unobserved model parameters, meaning that the target state can be estimated with minimal assumptions or prior knowledge regarding the object parameters. Results are validated against both a standard Unscented Kalman Filter (UKF) and Monte Carlo sampling

ORCID iDs

Parsonage, Benjamin ORCID logoORCID: https://orcid.org/0009-0001-3313-9661 and Vasile, Massimiliano ORCID logoORCID: https://orcid.org/0000-0001-8302-6465;
CORE (COnnecting REpositories)