To address the challenges in UAV aerial images, including the low pixel proportion of small objects, the attenuation of edge textures during downsampling, and interference from complex backgrounds, this study proposes CGSF-DETR, a real-time object detection algorithm for UAV aerial images. Through the GFBlock feature extraction module, spatial and frequency information is combined to effectively extract global contextual information, enhance the edge representation of small objects, and suppress background interference. In addition, an AIFI-LRSA encoding module is proposed, which uses pyramid pooling attention to replace traditional multi-head self-attention, thus enhancing long-range dependency modeling while keeping the computational cost largely unchanged. Furthermore, a P2 detection head is introduced and combined with the DGST fusion module to achieve effective complementarity between shallow details and deep semantics. Experimental results demonstrate that, compared with the original model, the proposed method improves mAP50 by 4.3%, 5.6%, and 1.4% on the VisDrone-2019, AI-TOD, and DOTAv1.0 datasets, respectively, while reducing the number of parameters by 32.5%.