Lightweight real-time detection network for fuel tank truck filling ports in complex scenarios

Abstract To improve filling ports detection accuracy and stability in complex autonomous fueling scenarios while maintaining real-time performance, this paper proposes an enhanced detection model based on YOLO11n, termed EFDS-YOLO. First, an efficient multi-scale feature fusion module is designed to strengthen the extraction of global, mid-level, and local features while maintaining model lightweightness. Second, a frequency-domain enhanced C3k2 module and a dynamic fine-grained channel attention mechanism are incorporated to further enhance the representation of key features. Third, the Shape-Intersection over Union loss function is introduced to improve the localization accuracy of fueling port through a more refined shape-matching strategy. On a self-constructed tanker truck filling port dataset, EFDS-YOLO achieves an average of 76.1% mAP@0.75, 6.3 G FLOPs, 2.5 M parameters, and 72.3 FPS across core metrics. Compared with YOLO11n and 19 other baseline models, it improves mAP@0.75 by an average of 2.1%, while reducing parameter count and computational cost by 7.4% and 4.5%, respectively. On a public fuel-cap-only dataset, EFDS-YOLO achieves an average of 98.5% precision and 98.9% mAP@0.5, outperforming YOLO11n and 19 other models by 3.8% and 2.2%, respectively. Moreover, real-time detection experiments conducted on a resource-constrained autonomous fueling-arm platform demonstrate that EFDS-YOLO maintains real-time capability as well as high detection accuracy and stability under various complex scenarios, further validating the effectiveness and practicality of the proposed model.