Absolute Visual Localization for Aerial Vehicles Based on Deep Learning Image Matching

The range of applications for aerial vehicles is expanding rapidly, encompassing various fields such as transportation, security, rescue, and mapping. The foundation for the execution of these tasks is accurate localization. The prevailing localization method for aerial vehicles is currently reliant on the Global Navigation Satellite System (GNSS). However, this method is vulnerable to disruptions caused by occlusion or interference, particularly in environments such as mountainous regions, forests, and urban canyons. Inertial navigation systems and visual odometers are commonly employed in GNSS-denied environments; however, their long-term accuracy is poor due to error accumulation. Absolute visual localization, based on image matching, exhibits superior long-term accuracy and does not require external devices, making it a suitable alternative when GNSS is unavailable. Additionally, deep learning has demonstrated remarkable potential in image matching, offering higher accuracy compared to traditional methods. The proposed methodology utilizes a deep learning-based image matching technique, employing the LoFTR model to match features on rectified aerial images and segmented satellite maps. Addressing the sensitivity of the conventional regional perspective transformation to nonlinear image differences, a central perspective transformation is proposed to calculate the precise position of the aerial vehicle on the map. The experimental results, obtained from flights at an altitude of 500 m, demonstrate that the proposed method, with a spatial resolution of 1.1 m for the aerial images and 2.4 m for the map blocks, exhibits a mean absolute error of 17.39 m. This outcome signifies a 54.55% reduction in error when contrasted with the conventional approach on the same dataset, attaining a level of accuracy comparable to that of general GNSS localization. Furthermore, the proposed method exhibits robustness to seasonal alterations in ground scene, providing an accurate and reliable solution for aerial vehicle localization in GNSS-denied environments.