This paper proposes an improved deep learning model, CGOD-YOLOv5, for precise marker detection to support autonomous UAV landing in complex environments. The model integrates ODConv and C3GAM into YOLOv5 to incorporate a global attention mechanism to enhance object detection accuracy without increasing network complexity. Evaluation is conducted on a custom marker dataset comprising 3200 UAV-captured or simulated images, as well as the CARLA urban simulation dataset. Experimental results show that CGOD-YOLOv5 achieves a of 96% on the marker dataset, an improvement of 0.8% over the baseline YOLOv5s. It also achieves 89.4% on the CARLA dataset, outperforming the original in both precision and generalization. The proposed method demonstrates robust detection of small, occluded, or low-resolution markers, making it suitable for real-time embedded UAV applications in scenarios such as environmental monitoring, precision agriculture, traffic management, and autonomous landing.