/UAVLocalization

Implementation of visual based UAV geo-localization using satellite imagery

Primary LanguagePython

Leveraging Map Retrieval and Alignment for Robust UAV Visual Geo-Localization

Introduction

This project focuses on development of a robust geo-localization system on aerial platform leveraging deep-learning based map retrieval and alignment. Two public datasets from Ageagle have been re-organized to evaluate the proposed algorithms. A field test in Beijing Haidian has been also conducted to demonstrate the effectiveness of the localization system.

Input data: orthophoto and the target referenced map

Output data: extracted geo-coordinates

Get-Started

Install dependencies:

The environment we use can be seen in setup/environment.yml. Note this project is mainly built based on the pytorch without many additional dependencies. And this environment list can be referred if there is any conflicts of dependencies.

Prepare the dataset:

We use the pre-stored images to represent the scenes captured during the flight.

  • Our datasets: please download the datasets and input in the dataset directory.
  • Custom dataset: please make sure the map contains the actual geo-coordinates and the file of the query images should be re-named as such format: @index@longitude@latitude@.

Test on the dataset:

Please make sure the paths for the pretrained weights and the datasets are correct. With the evaluation for the Ageagle dataset, simply run:

python main.py

If other datasets need to be tested, please change the configuration in utility/config.py.

Absolute-Localization-Flow

The fine localization is achieve with frame-to-map alignment. For more details, please refer to the main.py and files in scripts/.

Code-Structure

The file structure is shown as the following. At present, we only provide the main files, and all the related files will be released after the article is published.

.
+--- asset          # asset for this repository
+--- datasets       # path to save the geo-referenced map and captured frames
+--- models         # path to save the pretrained network weights
+--- scripts        # essential scripts for network models
+--- setup          # statement for the dependencies
+--- utility        # essential utilities to load image and visualize
+--- main.py        # main programme
+--- README.md

Resources-Download

  • DATASETS: All the datasets for this research have been open-sourced at the this link.

  • WEIGHTS: The model checkpoints have been also given at the this link.

(only the Ageagle datasets are available at present.)

Acknowledgements

In particular, we appreciate the following online resources to support the training and testing in this work.

We also express our gratitude for these open-sourced researches and parts of this work are inspired by them.

Citation

(related publication waiting for reviewing)