Visual-inertial navigation of UAVs with adaptive uncertainty estimation of neural network localization based on evidential learning

Visual-inertial navigation of UAVs with adaptive uncertainty estimation of neural network localization based on evidential learning

Lazareva P.A. (KAI, Kazan, Russia)
Gurko N.V. (KAI, Kazan, Russia)

Abstract

The problem of unmanned aerial vehicle navigation in GNSS-denied environments is considered. A hybrid system based on an error-state Kalman filter (ESKF) is proposed, integrating inertial measurements, optical-flow velocity estimation, and absolute visual localization against satellite maps using XFeat neural-network descriptors and the LighterGlue matching algorithm. To adaptively form the measurement covariance of the neural localization module, an evidential deep learning (EDL) approach with a Normal-Inverse-Gamma parameterization is proposed, decomposing localization uncertainty into aleatoric and epistemic components in a single forward pass. A notion of explainable observability is introduced as a context-dependent measure of absolute-measurement informativeness, a theorem on observability degeneracy under standard process covariance is proven, and a modification ensuring a non-zero Kalman gain between updates is given. A systematic ablation study over four measurement-covariance formulations, two process-covariance regimes, and an online strategy-switching algorithm was conducted on three flight datasets. The evidential covariance outperforms the heuristic on all three datasets; the gain grows monotonically with scene informativeness, reducing the mean trajectory error by 7–46 % relative to the heuristic. The magnitude of the gain is governed by the correlation between the EDL-predicted error and the actual error (ranging from 0.16 to 0.96 across datasets), which provides a quantitative applicability criterion. The clipped formulation, anchoring the estimate to the heuristic prior, is shown to be a necessary robustness mechanism on domains where the epistemic component of the NIG output is miscalibrated.

Keywords

unmanned aerial vehicle; visual-inertial navigation; error state Kalman filter; absolute visual localization; neural network descriptor; explainable artificial intelligence; deep learning; epistemic uncertainty.

Edition

Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 259-278

ISSN 2220-6426 (Online), ISSN 2079-8156 (Print).

DOI: 10.15514/ISPRAS-2026-38(6)-17

For citation

Lazareva P.A., Gurko N.V. Visual-inertial navigation of UAVs with adaptive uncertainty estimation of neural network localization based on evidential learning. Proceedings of the Institute for System Programming, vol. 38, issue 6, part 1, 2026, pp. 259-278 DOI: 10.15514/ISPRAS-2026-38(6)-17.

Full text of the paper in pdf (in Russian) Back to the contents of the volume