Abstract

As various unmanned aerial vehicles (UAVs) begin executing diverse missions over urban areas, they are transforming city logistics from two-dimensional ground transportation to three-dimensional aerial delivery. However, the gradient flow field variations around buildings and unpredictable air currents within dense urban canyons pose significant challenges for safe UAV flight. Therefore, a method capable of reconstructing the fluid-structure interaction (FSI) relationship within urban wind fields and analyzing their three-dimensional structure is a key foundation for advancing urban logistics from flat to three-dimensional operations. This study aims to achieve the informatization of flow field data in atmospheric boundary layer wind tunnels using deep learning-based reconstruction methods. Smoke visualization techniques were deployed to track airflow around buildings, while computer vision-based algorithms reconstructed wind fields from sparse multi-angle video recordings to investigate fluid-structure interactions between structures and flow fields. Furthermore, measurements from porous pressure probes validated the deep learning model’s predictions of velocity profiles and turbulence intensity under low wind speed conditions. Experimental results demonstrate that this approach successfully reconstructs high-resolution, time-series flow fields within complex, obstructed three-dimensional simulated urban environments using machine learning techniques.