Landmark-aware visual localization for GNSS-denied UAV wind farm inspection via lifecycle binding and adaptive fusion

Abstract This paper proposes a lightweight landmark-aware visual localization framework intended as a backup localization subsystem for UAV inspection in GNSS-denied wind farms. The system utilizes wind turbine towers as natural semantic beacons and integrates YOLOv12s and BoT-SORT for real-time detection and robust tracking. Absolute pose estimation is achieved via a bbox-based 2.5-D image–altimeter geometry model, a landmark binding lifecycle management strategy (LBL), and an adaptive spatio-temporal weighted fusion algorithm (ASTW). In five repeated RflySim trials, the combined framework achieved a three-dimensional localization RMSE of $$8.33\\pm 1.20$$ m compared with $$234.8\\pm 42.3$$ m for the baseline, with end-to-end latency of approximately 200 ms per processed frame. In field validation at a 26-turbine wind farm, the algorithm ran online on a Jetson Xavier NX and independently generated visual position estimates without using RTK/GNSS data, achieving a three-dimensional RMSE of 24.95 m and a latency of 220 ms per processed frame. Because RTK/GNSS remained active for waypoint flight control and was recorded as ground truth, this experiment validates online GNSS-independent visual localization rather than closed-loop autonomous navigation under a physical GNSS outage. Because the field evaluation comprised one mission at one wind farm under favorable weather, the field result supports feasibility for area-level localization and coarse navigation assistance, but not yet high-precision close-range turbine inspection or broad generalization across sites and weather conditions.