Autonomous navigation supported by a visual digital twin

Sammanfattning

Unmanned Aerial Systems (UAS) commonly rely on Global Navigation Satellite Systems (GNSS) for localization and autonomous navigation. However, GNSS signals are vulnerable to interference, jamming, and urban canyon effects, motivating the need for alternative localization methods. This thesis investigates whether a visual digital twin, consisting of georeferenced aerial imagery and a 3D point cloud, can support absolute drone localization in GNSS-denied environments. A map-based localization approach is proposed in which visual descriptors are extracted offline from aerial reference imagery and associated with georeferenced 3D points to form a descriptor-enhanced digital twin. During operation, descriptors extracted from incoming drone images are matched directly against this map, producing 2D–3D correspondences that are used to estimate the drone's global pose through Perspective-n-Point (PnP) and RANSAC-based geometric verification. By localizing each frame independently against the digital twin, the method avoids the cumulative drift associated with monocular visual odometry. The approach was evaluated using real drone imagery captured under substantial viewpoint and appearance differences from the aerial reference data. Across multiple experiments, pose estimation succeeded for 68,2–90,9\% of frames, achieving mean horizontal position errors between 12,0 and 29,6 m and mean vertical errors between 8,1 and 16,1 m. Higher image resolutions consistently improved localization accuracy, while larger sets of reference views provided more precise position estimates in areas with dense map coverage. The results demonstrate that direct matching against a descriptor-enhanced visual digital twin is a viable approach for drift-free global localization without GNSS. Although the achieved accuracy is not yet sufficient for standalone deployment, the findings indicate that visual digital twins can provide a promising foundation for robust navigation in GNSS-challenged environments and highlight key factors limiting current performance, including cross-view appearance differences and reference data quality.