Year: 2026
Type: preprint
Abstract: Bayesian state estimation for discrete-time nonlinear stochastic systems is considered in the presence of measurement outliers. Building on a fixed-structure Gaussian mixture filtering framework, this paper proposes a robust measurement-update variant in which the predictive density structure is determined by an offline decomposition of the transition density into axis-aligned Gaussian components. This construction maintains the Gaussian mixture structure as deterministic and tunable via the chosen decomposition fidelity. Measurement components affected by outliers are modeled using a Student's-t distribution, and the corresponding update of each Gaussian mixture component is approximated by a variational Bayes procedure. The resulting filter is evaluated in a three-dimensional tracking scenario with range and bearing measurements, where the bearing channel is affected by outliers, modeled as heavy-tailed noise.
Source: arXiv (Cornell University)
Authors: Straka, Ondřej, Hanebeck, Uwe D.
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203909920
Cites:
Cited by:
Related to: