Helipad target detection method for low-altitude rotor UAVs based on improved YOLOv11 network model

Target detection for precise helipad recognition and localization in low-altitude rotor UAVs is a critical research area. Due to the inconspicuous visual appearance of ground targets, deep learning vision models struggle to detect targets accurately in complex environments. To finalize our research efforts, this empirical study improves the spatial feature extraction of the YOLOv11 model’s C3K2 module by introducing the SPD-Conv algorithm, which converts spatial information into depth information to enhance feature extraction. Additionally, the PSA module in YOLOv11 suffers from limited global modeling capabilities. We mitigate this by incorporating the global attention mechanism, enabling the model to better capture contextual relationships and improve classification accuracy. Compared with traditional YOLO and YOLOv12 models, the improved YOLOv11 achieves a precision (PPV) of 97.3% and sensitivity (TPR) of 93.1%. Mean average precision (mAP) at IoU thresholds of 0.5 and 0.5–0.95 reaches 97.6% and 86.4%, respectively—improvements of 7.2, 3.7, 0.8, 3.8, and 7.3% points over model of YOLOv12. Simulations using UAV-based visual systems confirm significant performance gains in target detection accuracy.