Scholarlink

<p>Potholes pose significant risks to road safety, vehicle integrity, and transportation efficiency, while also increasing infrastructure maintenance costs. Conventional pothole detection approaches, such as manual inspections and sensor-based systems, are often costly, time-consuming, and impractical for large-scale deployment. This study proposes an integrated comparative framework for automated pothole detection that integrates image classification, real-time video-based potholes detection. Multiple convolutional neural network architectures including VGG16, VGG19, InceptionV3, MobileNetV2, DenseNet121, DenseNet201, EfficientNetB3, and a task-specific custom CNN are evaluated for pothole classification. The proposed custom CNN is specifically designed to capture pothole-related texture and shape characteristics and achieves a classification accuracy of 98.7%, demonstrating competitive performance with substantially reduced model size. Although EfficientNetB3 attains the highest classification accuracy of 99.54%, its larger architectural complexity increases computational and memory requirements. In contrast, the proposed custom CNN delivers near-equivalent accuracy with a significantly smaller model footprint and lower inference cost, making it more compatible with real-time and edge-based deployment scenarios. For real-time localization using bounding boxes, YOLOv8-based models are employed to detect potholes and generate predictions on road images and video streams. The complete pipeline encompasses dataset collection, preprocessing, annotation, and model training to ensure robust performance under diverse road and environmental conditions. Model performance is assessed using standard classification metrics (accuracy, precision, recall, and F1-score) and detection metrics (mean Average Precision), with YOLOv8 models achieving inference latency below 30 ms per frame, supporting real-time operation. The proposed framework offers a scalable and cost-effective solution for intelligent pothole detection, with potential applications in smart city infrastructure, intelligent transportation systems, and real-time road maintenance.</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

A lightweight deep learning framework for image-based pothole classification and real-time video-based pothole detection

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月30日

查看全文
  • Rizuanul Kaisar,
  • Kazi Naimur Rahman,
  • Homayun Kabir

摘要

Potholes pose significant risks to road safety, vehicle integrity, and transportation efficiency, while also increasing infrastructure maintenance costs. Conventional pothole detection approaches, such as manual inspections and sensor-based systems, are often costly, time-consuming, and impractical for large-scale deployment. This study proposes an integrated comparative framework for automated pothole detection that integrates image classification, real-time video-based potholes detection. Multiple convolutional neural network architectures including VGG16, VGG19, InceptionV3, MobileNetV2, DenseNet121, DenseNet201, EfficientNetB3, and a task-specific custom CNN are evaluated for pothole classification. The proposed custom CNN is specifically designed to capture pothole-related texture and shape characteristics and achieves a classification accuracy of 98.7%, demonstrating competitive performance with substantially reduced model size. Although EfficientNetB3 attains the highest classification accuracy of 99.54%, its larger architectural complexity increases computational and memory requirements. In contrast, the proposed custom CNN delivers near-equivalent accuracy with a significantly smaller model footprint and lower inference cost, making it more compatible with real-time and edge-based deployment scenarios. For real-time localization using bounding boxes, YOLOv8-based models are employed to detect potholes and generate predictions on road images and video streams. The complete pipeline encompasses dataset collection, preprocessing, annotation, and model training to ensure robust performance under diverse road and environmental conditions. Model performance is assessed using standard classification metrics (accuracy, precision, recall, and F1-score) and detection metrics (mean Average Precision), with YOLOv8 models achieving inference latency below 30 ms per frame, supporting real-time operation. The proposed framework offers a scalable and cost-effective solution for intelligent pothole detection, with potential applications in smart city infrastructure, intelligent transportation systems, and real-time road maintenance.