Zhen Wang Dianxi Shi Chunping Qiu Songchang Jin Tongyue Li Ziteng Qiao Yanyan Shi
Intelligent Game and Decision Lab, Academy of Military Sciences, China.
Abstract
In recent years, vision-based localization techniques have emerged as effective solutions for Unmanned Aerial Vehicle (UAV) localization in global navigation satellite system (GNSS) denied conditions. However, most UAV visual localization methods rely on high-precision satellite images or complex 3D maps, which, unfortunately, are not only costly to create, store, and maintain but also susceptible to seasonal variations. In contrast, humans can successfully determine their location using simple vectorized maps. Motivated by this human capability, we propose MapLocNet, the first deep neural network designed to localize UAVs by exploiting vectorized maps. MapLocNet is an end-to-end network that consists of two branches dedicated to extracting features from UAV images and vectorized maps. The features obtained from both branches are subsequently transformed into the frequency domain and multiplied to facilitate correlation analysis. An inverse Fourier transform is then applied to generate a correlation map. Finally, the peaks in the correlation map are detected to precisely determine the UAV's position. To evaluate MapLocNet, we introduce a comprehensive and challenging dataset encompassing seven cities worldwide. Through rigorous experimentation, MapLocNet has demonstrated competitive performance compared to existing methods. Furthermore, we have included a real-scene dataset to verify the generalization ability of the proposed model. The code and datasets can be accessed at https://map.geovisuallocalization.com.
Online Demo(coming soon)
Video
Download
|
|
Last update: Aug. 18, 2023
Copyright © 2023. All rights reserved.