Vision-Based Aerial Navigation Using Satellite Offline Maps

While Global Navigation Satellite System (GNSS) technology combined with advanced augmentation techniques provides high accuracy, its reliance on signal integrity leaves safety-critical applications vulnerable to spoofing and jamming. To address this, a lightweight vision-based localization module is proposed as a self-contained GNSS substitute for small Uncrewed Aerial Vehicles (UAVs). It leverages offline satellite maps and an ensemble of Convolutional Neural Network (CNN) models to estimate 2D offsets and associated uncertainty for stable EKF fusion. The system is designed for seamless integration with the PX4 Autopilot and real-time operation on resource-limited platforms. Software In The Loop (SITL) simulations over varied mission scenarios at Virginia Tech's Kentland Farm demonstrate an R95 accuracy of 4.5 m to 5.5 m, with effective drift bounding even in low-texture environments. Remaining challenges include uncertainty calibration, particularly over tree-covered regions, before transitioning to field testing.