RFP-YOLO26: A Fixed Single-Cycle Feedback Feature Pyramid for Slender Obstacle Detection in Autonomous Street-Sweeping Vehicles

Slender obstacles, such as ropes, cables, and rubber hoses, may interfere with the operation of autonomous street-sweeping vehicles because of their narrow shapes and weak visual features. This study presents RFP-YOLO26 as an applied detector-design and systems-integration approach that combines established SPDConv, C3k2_Faster_EMA, and SimAM modules with a fixed single-cycle feedback feature pyramid consisting of an initial top-down pass, one bottom-up feedback pass, and a second top-down refinement pass. Experiments were conducted on the proprietary USLO dataset using a random image-level split. On the current internal test set, RFP-YOLO26 achieved 97.9% mAP@0.5, 65.0% mAP@0.5:0.95, 98.1% precision, and 95.6% recall. Compared with YOLOv26n, these values represent increases of 2.9, 2.2, 1.7, and 4.3 percentage points, respectively. RFP-YOLO26 contains 4.21 M parameters and requires 9.9 GFLOPs, compared with 2.38 M parameters and 5.2 GFLOPs for YOLOv26n. Deployment on the Jetson Orin Nano indicates embedded execution feasibility under the reported configuration. The primary benchmark remains a seed-0 descriptive comparison. The supplementary five-seed analysis and the seed-0 principal-model comparison under route-disjoint Split A retained the same model ordering; the additional route-disjoint runs provided descriptive route-level summaries. However, the available evidence does not establish universal statistical superiority, external generalization, or improved operational safety.