​Abstract: Today there is a wide range of different navigation techniques with additional technologies paving the way for more complex fusion approaches. As the current standard, a combination of Inertial Navigation System (INS) and Global Navigation Satellite System (GNSS) enhanced with state-of-the-art augmentation services provide high accuracy, global coverage, and an acceptable baseline robustness. However, its reliance on signal integrity leaves safety-critical applications vulnerable to spoofing and jamming, while global crises increase the likelihood of such destructive interferences. With self-containment and high deployment volumes in mind, a lightweight vision-based reference module is proposed as a GNSS fallback for seamless integration into small Uncrewed Aerial Vehicles (UAVs). Using offline satellite maps and an ensemble of Convolutional Neural Network (CNN) models, 2D offsets and associated uncertainties are fused through the PX4 autopilot’s on-board Extended Kalman Filter (EKF). Software-In-The-Loop (SITL) simulations provide first results on the feasibility and real-time capabilities of the module on a resource-limited platform. Systematic mission parameter variation on terrain data of Virginia Tech’s Kentland Farm demonstrates the required minimum flight path stabilization capabilities with an R95 accuracy of 4.5 m to 5.5 m. Even in low-texture environments drift stays bounded, while varying signal loss scenarios reveal current stability limitations. Driven by the integration concept and initial SITL results, the resulting hardware test platform is presented as the foundation for upcoming flight testing.

Keywords: GNSS-denied navigation, Vision-based sensing, Cross-view geolocalization.

__________________________________________________________________________________________