Mesh-Map-Assisted Monocular 3D Localization for UAVs: A Multi-Scale Approach

Abstract

Unmanned aerial vehicles (UAVs) are increasingly utilized across military, civilian, and agricultural sectors, necessitating accurate and efficient 3D target localization. Traditional 2D detectors lack depth perception, while stereo vision and LiDAR have range-dependent and hardware limitations, respectively. To address these limitations, we propose a monocular 3D localization framework using a prebuilt, georegistered triangular mesh. A two-dimensional grid indexes the mesh, and a progressive multi-scale strategy loads high-resolution blocks for ray–mesh intersection. The nearest intersection is used only for targets supported by, or projectable onto, the mapped surface. Unity experiments give localization RMSEs of 0.0193–0.9229 m; a real UAV field log gives a pooled static-track repeatability RMSE of 0.0849 m. Repeated timing measurements give mean per-query times of 5.08 ms and 1.03 ms for full-resolution and progressive search, respectively.