VecMapLocNet: Vision-based UAV localization using vector maps in GNSS-denied environments

Abstract

Vision-based localization techniques are effective UAV localization solutions for GNSS-denied conditions, however they depend on costly, complex, and seasonally variable satellite images or 3D maps, whereas humans can determine location using vector maps. Inspired by human navigation strategies, we propose VecMapLocNet , which uses vector maps to determine UAV 3-DoF poses (latitude, longitude, and yaw) through cross-modal matching. Three key modules are designed to improve the matching between UAV images and vector maps. The UAV feature extraction module is low-latency and adaptable to various flight altitudes, ensuring it is suitable for airborne deployment. The vector map feature extraction module employs a weighted representation of different map elements, ensuring robustness against changes in visual appearance. Inspired by Fourier transforms, the feature matching module for 3-DoF pose estimation is parameter-free, computationally efficient, and invariant to cross-modal differences. To evaluate VecMapLocNet, we introduce a comprehensive dataset that presents challenges, encompassing seven cities worldwide. Through rigorous experimentation, VecMapLocNet has demonstrated competitive performance compared to existing methods in localization accuracy (84.45% Recall@5 m), yaw estimation (88.61% Recall@5°), and computational efficiency (25.23ms latency on onboard device Jetson Orin). Furthermore, we validated VecMapLocNet’s performance in real-world scenarios, with experimental results confirming its generalization ability, achieving a localization error of 16.7 m and an orientation error of 3.1°. The code and datasets are available at https://map.geovisuallocalization.com.