Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Leveraging edge detection and neural networks for better UAV localization

Theo Di Piazza, Enric Meinhardt-Llopis, Gabriele Facciolo, Benedicte Bascle, Corentin Abgrall, Jean-Clement Devaux

TL;DR

This work tackles GNSS-denied UAV geolocalization by converting UAV views into edge maps and encoding them into a latent embedding of dimension $1024$ for tiles of size $256\\times256$, then matching against a georeferenced reference set. Localization is performed via embedding similarity (dot-product) with a Lowe's ratio-based confidence check (threshold $1.13$) to filter uncertain predictions. The study compares BoVW, Auto-Encoder, and Triplet methods, finding that edge-preprocessed embeddings—especially the Auto-Encoder—significantly improve accuracy and robustness to illumination, rotation, and altitude perturbations. The proposed approach enables reliable, real-time onboard localization in GNSS-denied scenarios and informs edge-based preprocessing for geo-referenced image matching.

Abstract

We propose a novel method for geolocalizing Unmanned Aerial Vehicles (UAVs) in environments lacking Global Navigation Satellite Systems (GNSS). Current state-of-the-art techniques employ an offline-trained encoder to generate a vector representation (embedding) of the UAV's current view, which is then compared with pre-computed embeddings of geo-referenced images to determine the UAV's position. Here, we demonstrate that the performance of these methods can be significantly enhanced by preprocessing the images to extract their edges, which exhibit robustness to seasonal and illumination variations. Furthermore, we establish that utilizing edges enhances resilience to orientation and altitude inaccuracies. Additionally, we introduce a confidence criterion for localization. Our findings are substantiated through synthetic experiments.

Leveraging edge detection and neural networks for better UAV localization

TL;DR

This work tackles GNSS-denied UAV geolocalization by converting UAV views into edge maps and encoding them into a latent embedding of dimension for tiles of size , then matching against a georeferenced reference set. Localization is performed via embedding similarity (dot-product) with a Lowe's ratio-based confidence check (threshold ) to filter uncertain predictions. The study compares BoVW, Auto-Encoder, and Triplet methods, finding that edge-preprocessed embeddings—especially the Auto-Encoder—significantly improve accuracy and robustness to illumination, rotation, and altitude perturbations. The proposed approach enables reliable, real-time onboard localization in GNSS-denied scenarios and informs edge-based preprocessing for geo-referenced image matching.

Abstract

We propose a novel method for geolocalizing Unmanned Aerial Vehicles (UAVs) in environments lacking Global Navigation Satellite Systems (GNSS). Current state-of-the-art techniques employ an offline-trained encoder to generate a vector representation (embedding) of the UAV's current view, which is then compared with pre-computed embeddings of geo-referenced images to determine the UAV's position. Here, we demonstrate that the performance of these methods can be significantly enhanced by preprocessing the images to extract their edges, which exhibit robustness to seasonal and illumination variations. Furthermore, we establish that utilizing edges enhances resilience to orientation and altitude inaccuracies. Additionally, we introduce a confidence criterion for localization. Our findings are substantiated through synthetic experiments.
Paper Structure (6 sections, 3 equations, 3 figures, 3 tables)

This paper contains 6 sections, 3 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Dataset
  4. Methodology
  5. Results analysis
  6. Conclusion

Figures (3)

  • Figure 1: Overview of the method for edge extraction and Auto-Encoder. Offline, the Auto-Encoder is trained to reconstruct edges of the tiles from the reference area. Then embeddings of reference tiles and weights of the encoder are loaded on board the UAV. During the UAV flight, edges are extracted from the UAV view before being given to the encoder which produces an embedding. Last, the UAV's position is estimated by comparing the embedding of the UAV's view with the embeddings of the reference tiles.
  • Figure 2: Illustration of the process for generating the reference tiles from imagery of the reference area.
  • Figure 3: Sensitivity to rotation and altitude drop for Auto-Encoder and Triplet Model, on trajectory B. For the row Orientation, figures show location accuracy at 15m VS angle of variation (degrees). For the row Altitude, figures show location accuracy at 15m VS altitude drop (meters).