Joint 3D modeling and motion estimation of dynamic wind-turbine blades from monocular UAV videos

Abstract

Non-stop inspection of wind turbines has gained significant attention due to its potential to minimize downtime and reduce maintenance costs. While Unmanned Aerial Vehicles (UAVs) have been successfully deployed for surface defect detection on rotating blades, 3D modeling of operational blades remains largely unexplored. Such modeling offers immense potential for analyzing blade geometric behaviors under aerodynamic wind loads. To address this gap, we propose a monocular UAV-based framework that adapts dynamic Neural Radiance Fields (NeRF) for the inspection and 3D reconstruction of in-service wind turbine blades. Our approach uniquely optimizes blade’s 3D geometry, texture, and spatial motion using only a single UAV-mounted camera, by integrating optical-flow to constrain 3D motion and utilizes pseudo-depth supervision to accelerate the modeling process, which distinguish from most dynamic frameworks that require synchronized multi-view camera arrays. Extensive experiments across various real-world conditions demonstrate that our method enables high-fidelity dynamic modeling (PSNR ' 28 and motion MAE in only 0.05 pixel), marking a significant step toward the dynamic 3D monitoring of operational wind turbine blades.