GNSS-denied visual localization ameliorative method for UAVs in non-urban environments

Abstract

In the context of Unmanned Aerial Vehicles (UAVs), localization is critical for both military and civilian applications. This is particularly true in environments without urban infrastructure, where Global Navigation Satellite System (GNSS) signals are unavailable. In these settings, vision-based methods have emerged as a promising solution. Despite their potential, current deep learning-based matching algorithms exhibit significant limitations in accurately localizing UAVs. To address this, our paper introduces enhanced algorithms that build upon existing methods. Specifically, we propose the use of the DC-ShadowNet shadow removal algorithm for UAV image preprocessing, a critical step in urban areas where shadows from large structures can obscure ground details, especially under sunny conditions. Additionally, we employ an improved matching algorithm based on the ASpanFormer model to increase accuracy in image matching. Our testing shows that these advancements lead to improved localization accuracy, both on a public dataset and on actual flight data. Furthermore, our method is well-suited for long-duration flights and offers considerable advantages in urban environments when compared to previous state-of-the-art Visual Odometry techniques.