Detecting objects in real time is essential for unmanned aerial vehicle (UAV) applications such as monitoring, rescue operations, and traffic analysis. However, aerial imagery remains difficult to process because many objects are small, densely grouped, and visually degraded by clutter and perspective changes. Convolutional detectors like You Only Look Once (YOLO) capture local structures efficiently, but provide limited global reasoning, while pure Transformer models are often too heavy for embedded platforms. In this paper, we propose YOLOv11-WTB, which enhances YOLOv11 by inserting a single Window-Based Transformer Block (WTB) into the P3 stage of the backbone, where most small-scale details reside. The block performs multi-head self-attention within fixed, non-overlapping windows, improving feature context without redesigning the neck or the detection heads. On the VisDrone-2019 dataset, YOLOv11n-WTB increases mean Average Precision (mAP) at 0.5 Intersection over Union (IoU) from 46.1% to 51.6% and mAP@0.5:0.95 from 28.4% to 32.0%, while maintaining 66 frames per second (FPS). YOLOv11s-WTB reaches 52.6% mAP@0.5 and 31.1% mAP@0.5:0.95 at 39 FPS, outperforming YOLOv8s in overall accuracy and in crowded categories. These results show that strengthening only the high-resolution stage provides a practical way to boost UAV detection performance without sacrificing real-time execution.