An improved object detection framework based on YOLOv11 for UAV

Unmanned Aerial Vehicle (UAV) object detection presents significant challenges due to the prevalence of small targets, dense spatial distributions, and complex background interference. Conventional detection frameworks often struggle with insufficient feature representation and high false detection rates. To address these limitations, we propose a lightweight detection framework based on YOLOv11, incorporating three key improvements: (1) a bidirectional weighted feature pyramid network (BiFPN++) for enhanced multi-scale feature fusion, (2) high-resolution prediction heads and tiny detection branches to strengthen small-object perception, and (3) a dynamic sample assignment mechanism with adaptive loss weighting to optimize training supervision. The framework integrates structural re-parameterization and depthwise separable convolutions, achieving a 65% reduction in parameters and compressing the model to 18.7M while maintaining high accuracy. On the VisDrone benchmark, our method achieves 41.5% mAP@[0.5:0.95] and 22.7% AP_small, improving small-object recall by 22% and reducing false detections by 37.2% compared with the baseline. These results demonstrate that the proposed approach achieves a favorable balance between detection accuracy, efficiency, and robustness, making it well-suited for real-time UAV inspection and monitoring tasks.