Scholarlink

<p>Potholes are one of the major concerns in terms of infrastructure and safety. They are a major concern for the safety of bikers, pedestrians, and vehicles. For the safety and maintenance of roads, it is imperative that pothole detection is carried out in real time with high accuracy, especially in varying lighting and environmental conditions. This paper presents a proactive approach for pothole detection. The approach is based on edge computing, geospatial intelligence, and deep learning-based object detection. The performance of three fundamentally different approaches, namely YOLOv8n (You Only Look Once version 8, nano variant), YOLOv10n (You Only Look Once version 10, nano variant), and Faster Region-Based Convolutional Neural Network (Faster R-CNN), is evaluated by the proposed framework. The models were trained on a heavily augmented dataset comprising 15,335 images. A comprehensive data augmentation strategy that included geometric and photometric transformations, as well as weather simulations specific to pothole detection, was adopted to increase the initial dataset consisting of 2,475 images. The majority of effective model is installed on an OpenCV AI Kit - Depth (OAK-D) camera with a Raspberry Pi computing board, which shows the ability to perform the proposed method on an edge device with limited resources. Moreover, a new composite metric named Pothole Severity Index (PSI) is proposed, which allows municipal agencies to prioritize their pothole repairs, and pairs the output of the visual detection with real world geospatial data from OpenStreetMap (OSM). The performance of the suggested framework is confirmed by thorough statistical validation, which includes paired significance testing and 5-fold cross-validation. While the Non-Maximum Suppression Free (NMS-free) YOLOv10n produces similar results with slightly lower computational overhead, YOLOv8n achieves a mAP@0.5 of 0.9686 and a Mean Average Precision at Intersection over Union (IoU) threshold 0.5 (mAP@0.5:0.95) of 0.9185, making both appropriate for edge Internet of Things (IoT) deployment.</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

Deep learning and edge computing framework for proactive pothole detection with Internet of Things

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月3日

查看全文
  • Nalina V,
  • Jayarekha P,
  • Anitha H M,
  • Adarsh V,
  • Amit Chandrashekhar Hegde,
  • Deekshitha R,
  • Mukund Rao

摘要

Potholes are one of the major concerns in terms of infrastructure and safety. They are a major concern for the safety of bikers, pedestrians, and vehicles. For the safety and maintenance of roads, it is imperative that pothole detection is carried out in real time with high accuracy, especially in varying lighting and environmental conditions. This paper presents a proactive approach for pothole detection. The approach is based on edge computing, geospatial intelligence, and deep learning-based object detection. The performance of three fundamentally different approaches, namely YOLOv8n (You Only Look Once version 8, nano variant), YOLOv10n (You Only Look Once version 10, nano variant), and Faster Region-Based Convolutional Neural Network (Faster R-CNN), is evaluated by the proposed framework. The models were trained on a heavily augmented dataset comprising 15,335 images. A comprehensive data augmentation strategy that included geometric and photometric transformations, as well as weather simulations specific to pothole detection, was adopted to increase the initial dataset consisting of 2,475 images. The majority of effective model is installed on an OpenCV AI Kit - Depth (OAK-D) camera with a Raspberry Pi computing board, which shows the ability to perform the proposed method on an edge device with limited resources. Moreover, a new composite metric named Pothole Severity Index (PSI) is proposed, which allows municipal agencies to prioritize their pothole repairs, and pairs the output of the visual detection with real world geospatial data from OpenStreetMap (OSM). The performance of the suggested framework is confirmed by thorough statistical validation, which includes paired significance testing and 5-fold cross-validation. While the Non-Maximum Suppression Free (NMS-free) YOLOv10n produces similar results with slightly lower computational overhead, YOLOv8n achieves a mAP@0.5 of 0.9686 and a Mean Average Precision at Intersection over Union (IoU) threshold 0.5 (mAP@0.5:0.95) of 0.9185, making both appropriate for edge Internet of Things (IoT) deployment.