Scholarlink

<p>With the rapid advancement of smart cities, ensuring the safety of urban infrastructure through automated manhole cover monitoring has become increasingly important. This study proposes a lightweight detection framework tailored for edge deployment. A Shape-Aware Regularization loss is introduced into YOLOv12, embedding geometric priors via Huber loss to enhance localization accuracy. Furthermore, a compact variant, YOLOv12n-tiny, is developed through compound scaling, achieving a model size of only 5.18 MB. Experimental results show state-of-the-art performance, with 0.930 mAP50, and an inference latency of 3.38 ms. Compared with YOLOv12-n, the proposed model improves mAP50 by 1.6% and speed by 7%, while surpassing YOLOv11n by 1.24% in accuracy and 27% in efficiency. Integrated with the Segment Anything Model (SAM), the system enables automated risk assessment of damaged covers. Deployed on smart inspection vehicles, it offers a practical and low-cost solution for real-time infrastructure monitoring, thereby contributing to safer and more sustainable smart cities.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

An improved YOLO model for manhole cover defect detection and risk assessment

  • 期刊论文

  • 发表日期：

    2026年1月14日

查看全文
  • Yajun Liu,
  • Ruohua Zhou,
  • Jianfang Zhang,
  • Nan Sun

摘要

With the rapid advancement of smart cities, ensuring the safety of urban infrastructure through automated manhole cover monitoring has become increasingly important. This study proposes a lightweight detection framework tailored for edge deployment. A Shape-Aware Regularization loss is introduced into YOLOv12, embedding geometric priors via Huber loss to enhance localization accuracy. Furthermore, a compact variant, YOLOv12n-tiny, is developed through compound scaling, achieving a model size of only 5.18 MB. Experimental results show state-of-the-art performance, with 0.930 mAP50, and an inference latency of 3.38 ms. Compared with YOLOv12-n, the proposed model improves mAP50 by 1.6% and speed by 7%, while surpassing YOLOv11n by 1.24% in accuracy and 27% in efficiency. Integrated with the Segment Anything Model (SAM), the system enables automated risk assessment of damaged covers. Deployed on smart inspection vehicles, it offers a practical and low-cost solution for real-time infrastructure monitoring, thereby contributing to safer and more sustainable smart cities.