Unmanned aerial vehicle visual localization method based on deep feature orthorectification matching

In the context of satellite denial conditions, the foundation for the safe and reliable completion of various tasks by unmanned aerial vehicles (UAVs) is the acquisition of high-precision positioning information. Traditional image matching methods face challenges in guaranteeing security, exhibit poor positioning accuracy, and involve numerous matching constraints. Therefore, a visual positioning method based on deep feature orthorectification matching is proposed, which utilizes a deep learning network to extract depth features from orthorectified UAV aerial images and commercial maps, establishes matching relationships and subsequently calculates high-precision UAV position information. The impact of different factors on visual positioning accuracy is analyzed according to the visual measurement model, and offline experiments ae conducted using a dataset of hollow aerial images. The experimental results demonstrate that, compared with the traditional template matching methods based on histogram of oriented gradients (HOG) features, the proposed method improves positioning accuracy by 25%, and the positioning root mean square error (RMSE) is better than 15 m+0.5%H (for height below 5000 m), which shows certain engineering application value. © 2024 Editorial Department of Journal of Chinese Inertial Technology. All rights reserved.