<p>Small object detection in UAV imagery remains a significant challenge due to the loss of high-frequency details, extreme scale variations, severe occlusion, and the computational constraints of embedded platforms. Existing methods often fail to balance accuracy and efficiency under these conditions. To address this, we propose GCS-DETR, a global context-driven real-time detector with occlusion suppression. Firstly, we design a lightweight frequency-domain dynamic backbone network (FreqDyNet), which employs a cross-stage partial dynamic filter module to perform adaptive filtering in the frequency domain, enhancing the capture of small object details while significantly reducing computational redundancy. Secondly, we propose occlusion-aware multi-scale feature pyramid network (OAM-FPN), an occlusion-aware multi-scale feature fusion network combining wavelet transformation, adaptive multi-branch fusion, and hybrid attention mechanism. Finally, we introduce a loss function fusing normalized Wasserstein distance and multi-point distance (NWD-MPDIoU) to improve small object localization accuracy and training stability. Benefiting from the above designs, GCS-DETR achieves 3.0% and 3.5% mAP₅₀ improvements over RT-DETR on VisDrone2019 and HIT-UAV datasets, together with a 20.6% reduction in model parameters. Its successful deployment on the embedded platform Jetson Orin Nano Super confirms strong practical utility for reliable, real-time detection in resource-limited UAV applications.</p>