Abstract:To address the limitations of traditional detection models caused by image blur, increased noise, and low pixel occupancy of pedestrians in high-resolution UAV aerial images, a pedestrian recognition algorithm for aerial photography is studied. Collaborative optimization is performed on the detection model architecture and training strategies. Based on the improved PP-YOLOE model, a Transformer encoder is embedded to construct a long-range dependency-aware network for enhancing semantic decoupling in complex backgrounds, and global positional encoding is introduced to optimize spatial localization. A dataset slicing strategy based on SAHI sliding windows is adopted, combined with multi-scale dynamic augmentation methods such as CLAHE adaptive histogram equalization and random saturation perturbation, to construct a training dataset that balances small target feature preservation and sample diversity. Experimental tests on the VisDrone2019 dataset demonstrate that the method effectively identifies pedestrian targets under various detection environments, achieving higher detection accuracy and robustness compared to mainstream algorithms. Practical application verification indicates that the algorithm meets the requirements for high-precision and robust pedestrian detection in intelligent UAV inspection scenarios.
