Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Season-invariant GNSS-denied visual localization for UAVs

Jouko Kinnari, Francesco Verdoja, Ville Kyrki

TL;DR

The paper tackles GNSS-denied UAV localization under significant seasonal appearance changes by introducing a season-invariant image-to-map matching approach. It uses a Siamese CNN to compute a similarity score between UAV-captured patches and corresponding orthophoto map patches, and integrates this score into Monte Carlo Localization with a KDE-based confidence weighting. Key contributions include (i) a season-invariant CNN-based matching mechanism, (ii) a Gaussian kernel density estimation framework to quantify pose-hypothesis confidence, and (iii) demonstrations on both simulated orthoimage data and real UAV flights showing faster convergence and reduced localization error compared with six baselines. The approach enables robust, near real-time GNSS-denied localization across seasons and varying viewpoints, advancing practical UAV autonomy in challenging environments.

Abstract

Localization without Global Navigation Satellite Systems (GNSS) is a critical functionality in autonomous operations of unmanned aerial vehicles (UAVs). Vision-based localization on a known map can be an effective solution, but it is burdened by two main problems: places have different appearance depending on weather and season, and the perspective discrepancy between the UAV camera image and the map make matching hard. In this work, we propose a localization solution relying on matching of UAV camera images to georeferenced orthophotos with a trained convolutional neural network model that is invariant to significant seasonal appearance difference (winter-summer) between the camera image and map. We compare the convergence speed and localization accuracy of our solution to six reference methods. The results show major improvements with respect to reference methods, especially under high seasonal variation. We finally demonstrate the ability of the method to successfully localize a real UAV, showing that the proposed method is robust to perspective changes.

Season-invariant GNSS-denied visual localization for UAVs

TL;DR

The paper tackles GNSS-denied UAV localization under significant seasonal appearance changes by introducing a season-invariant image-to-map matching approach. It uses a Siamese CNN to compute a similarity score between UAV-captured patches and corresponding orthophoto map patches, and integrates this score into Monte Carlo Localization with a KDE-based confidence weighting. Key contributions include (i) a season-invariant CNN-based matching mechanism, (ii) a Gaussian kernel density estimation framework to quantify pose-hypothesis confidence, and (iii) demonstrations on both simulated orthoimage data and real UAV flights showing faster convergence and reduced localization error compared with six baselines. The approach enables robust, near real-time GNSS-denied localization across seasons and varying viewpoints, advancing practical UAV autonomy in challenging environments.

Abstract

Localization without Global Navigation Satellite Systems (GNSS) is a critical functionality in autonomous operations of unmanned aerial vehicles (UAVs). Vision-based localization on a known map can be an effective solution, but it is burdened by two main problems: places have different appearance depending on weather and season, and the perspective discrepancy between the UAV camera image and the map make matching hard. In this work, we propose a localization solution relying on matching of UAV camera images to georeferenced orthophotos with a trained convolutional neural network model that is invariant to significant seasonal appearance difference (winter-summer) between the camera image and map. We compare the convergence speed and localization accuracy of our solution to six reference methods. The results show major improvements with respect to reference methods, especially under high seasonal variation. We finally demonstrate the ability of the method to successfully localize a real UAV, showing that the proposed method is robust to perspective changes.

Paper Structure

This paper contains 15 sections, 3 equations, 9 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related work
  3. Methodology
  4. uav Localization
  5. Similarity scoring function structure
  6. Data
  7. Computing importance factor from similarity measure
  8. Experiments
  9. Experimental setting
  10. Prior on initial pose
  11. Odometry noise
  12. Localization under seasonal appearance variation
  13. Localization on real uav data
  14. Discussion
  15. Conclusions

Figures (9)

  • Figure 1: We develop a similarity scoring method for uav localization that is invariant to seasonal appearance change.
  • Figure 2: Model structure for the image similarity network $c = f(I_{\mathcal{M},X},I_{t})$. Branches share parameters.
  • Figure 3: Histograms of similarity measures and estimated probability densities.
  • Figure 4: Orthoimages used in simulated experiments as map (a), "summer" (b), and "winter" (c) measurements. Each orthoimage is 4800$\times$2987 m at 1 m/pixel resolution.
  • Figure 5: Errors in simulated localization experiments when using different similarity measures, over different types of terrain. Minor appearance change refers to summer-to-summer matching, while significant refers to winter-to-summer matching. Medians of mean errors after 20 and 80 updates annotated.
  • ...and 4 more figures