To enhance the operational safety of drone systems, the deployment of real-time surveillance technologies is essential. Recently monitoring frameworks increasingly rely on high-resolution cameras and CNN-based object detection algorithms. However, directly applying open-source detection models in real-world environments often results in missed detections and false positives, potentially compromising the reliability of drone operations. This study proposes an improved YOLOv8-based drone detection model specifically optimized for aerial surveillance tasks. The proposed architecture integrates depthwise large-kernel convolutions into the backbone to expand the receptive field, enabling better recognition of global contours in low-texture environments such as open skies. In addition, a BiFPN neck structure is adopted to strengthen multi-scale feature fusion, and a P2 branch is introduced to preserve high-resolution spatial cues from early layers. The model’s performance is evaluated by using YOLOv8-s as a baseline under standardized training conditions. Experiments are conducted with 50 training epochs, an input resolution of 640 × 640, and a batch size of 16. Quantitative evaluation using metrics such as precision, recall, and mAP demonstrates consistent improvements across all indicators. Notably, the proposed model exhibits enhanced detection capability for small and distant drone targets, confirming its effectiveness in real-world surveillance scenarios. Overall, the results verify that combining large-kernel receptive field enhancement with adaptive multi-scale fusion using BiFPN significantly improves detection accuracy and robustness in drone monitoring systems.