Research on Real-Time Manhole Cover Detection from Vehicle Based on Deep Learning

Real-time manhole cover detection with on-board sensors provides critical support for refined road maintenance. To address the limitations of existing methods regarding real-time performance and portability, this paper proposes a lightweight dual-branch framework CLF-YOLOv8 based on an improved YOLOv8 object detector. This model incorporates CloFormer as the backbone and leverages its local attention mechanism to enhance the representation of manhole covers. Meanwhile, Wise-IoU Loss is adopted instead of CIoU Loss to improve model stability and generalization. Experiments demonstrate that the proposed approach achieves 5.92% higher detection accuracy and 5.63% higher recall compared to the original YOLOv8, with 1.52 times faster detection speed. It also outperforms RT-DETR, YOLOv7, YOLOv5 in precision, speed and mAP. Results indicate the proposed technique significantly advances detection accuracy and real-time performance over existing methods. It holds broad application prospects in buried pipeline inspection, road maintenance, etc.