UVD-YOLO: an engineering vehicle detection algorithm for UAV-based inspection of oil and gas pipelines

UAVs have become essential for routine inspections to ensure the safe operation of oil and gas pipelines. Nevertheless, aerial top-view scenes feature small-sized engineering vehicles cluttered with complex backgrounds, resulting in low detection accuracy. To enhance automatic detection accuracy of engineering vehicles in UAV-based pipeline inspections, this study proposes the Unmanned Aerial Vehicle Pipeline Inspection Engineering Vehicle Detection YOLO(UVD-YOLO), an improved algorithm based on YOLOv 11(You Only Look Once, YOLO). First, a vehicle-aware contextual fusion module is designed to enhance small-target feature extraction via parallel multi-scale convolutions. Second, a noise-suppressed attention module employs grouped features and cross-spatial learning to focus on vehicle regions and suppress background interference. Third, a hierarchical vehicle feature fusion module adaptively combines low-level details with high-level semantics to improve multi-scale detection. Evaluated on a self-built pipeline inspection vehicle dataset, UVD-YOLO outperforms the baseline YOLOv 11-s by 4.5% in mAP@0.50 and by 3.5% in mAP@0.50 ~ 0.95. Compared to mainstream detection algorithms, it delivers superior accuracy while maintaining fast inference, demonstrating notable advantages in small-target detection and robustness against cluttered backgrounds. UVD-YOLO effectively addresses two major challenges in UAV-based pipeline inspection: detecting small engineering vehicle targets and mitigating background interference. It offers a robust technical solution for intelligent pipeline inspection and serves as a modular design reference for small-target detection in comparable scenarios.