Year: 2026
Type: article
Abstract: Aerial-to-satellite visual localization enables GNSS-denied UAV navigation, but the appearance gap between low-altitude (50–150 m) UAV imagery and nadir satellite tiles makes per-frame visual place recognition (VPR) unreliable. Under perceptual aliasing, high similarity matches are often geographically inconsistent, so naïve anchoring fails. We introduce NaviLoc, a training-free three-stage trajectory-level estimator that treats VPR as a noisy measurement source and exploits visual-inertial odometry (VIO) as a relative-motion prior. Stage 1 (Global Align) estimates a global SE(2) transform by maximizing an explicit trajectory-level similarity objective. Stage 2 (Refinement) performs sliding-window bounded weighted Procrustes updates. Stage 3 (Smoothing) computes a strictly convex MAP trajectory estimate that fuses VIO displacements with VPR anchors while clamping detected outliers. On a challenging low-altitude rural UAV benchmark, NaviLoc attains 19.5 m mean localization error (MLE) – a 16.0x reduction compared to state-of-the-art localization method AnyLoc-VLAD, and 32.1x compared to raw VIO drift. End-to-end inference runs at 9 FPS on Raspberry Pi 5, enabling real-time embedded deployment.
Source: Drones
Authors: Pavel Shpagin, Taras Panchenko
Institution: Taras Shevchenko National University of Kyiv
Language: English
Subfield: Aerospace Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $3,137
OpenAlex ID: W7126039171