Visual localization is a key capability for UAVs operating where GNSS signals are unavailable. A standard technique matches aerial images against geo-tagged orthographic maps and then estimates the camera pose from that correspondence. However, many such systems depend on a digital surface model (DSM) to relate image features to ground geometry.
The new paper, posted on arXiv, proposes a DSM-free alternative. The method is built around latent-depth variable projection and structure-adaptive coupling, which together aim to estimate pose directly from the map and image data without needing a separate surface model. Removing the DSM requirement could make the approach easier to deploy in areas where such models are missing or outdated.
The abstract does not include experimental details or quantitative results, so the practical performance of the method is not yet clear from this summary. Still, the work points to a possible simplification of map-based UAV localization in GNSS-denied environments.