Near-Real-Time Pixel-Wise 3D Position Estimation from UAV Images Using a Reference Model for Urban Monitoring

UAV-based monitoring has been increasingly employed in various urban applications, including traffic analysis and infrastructure inspection. For continuous monitoring, timely information must be derived from newly acquired UAV images while maintaining spatial consistency across observations; however, existing methods remain limited in satisfying these requirements for a single image. Therefore, the objective of this study is to rapidly estimate pixel-wise 3D positions in a common coordinate system from a single newly acquired UAV image. For this purpose, this study proposed a method using a pre-built reference model. It estimates the interior and exterior orientation parameters of the input image and reprojects the dense point cloud of the reference model to determine pixel-wise 3D positions. Its performance was evaluated under Light, Season, Sensor, and Combined scenarios representing varying acquisition conditions. Across all scenarios, the maximum difference in 3D RMSE relative to the reference model was 0.036 m, and the average processing time was below 3 s per image. The estimated 3D positions correspond to the scene geometry represented in the reference model because the method relies on this model. These results demonstrate rapid pixel-wise 3D position estimation with accuracy comparable to that of the reference model, supporting continuous urban monitoring.