Pedestrian detection in aerial eye view exhibits poor real-time detection performance in non-terrestrial network (NTN) environments, especially when using unmanned aerial vehicles (UAVs) for aerial surveillance. For adapting to the air-to-ground recognition task, we optimized the AG-ReID dataset by adding new images from UAVs and named it AG-ReID-Pro. In this study, we adopt Yolov5 as the base model and perform model optimization and hardware porting for resource-constrained edge devices. Specifically, we apply MobileNetV3 as the backbone network for feature extraction to reduce the parameters of the model. The optimized model is successfully deployed on K230 SoC, which achieves efficient operation using its built-in KPU and 2D acceleration engine. Experimental results show that the model achieves fast inference speed while maintaining high detection accuracy, with excellent real-time performance on the K230 board.