Scholarlink

<p>Reliable detection of road surface defects is important for traffic safety, pavement maintenance, and cost-effective infrastructure management, especially where repeated manual inspection is expensive. This paper presents a low-cost multi-sensor framework for automated and geo-localized road assessment using 2D LiDAR, a monocular camera, GPS, and an inertial measurement unit. LiDAR scans are converted into road-surface profiles, non-road structures are removed using density-based clustering and principal component analysis, and the low-frequency road undulation is separated from the raw elevation profile to extract high-frequency roughness signatures. Three detection strategies are evaluated: a filter-based roughness detector using autocorrelation, linear predictive coding, prediction power error, and roughness index; a Random Forest classifier for pothole and crack detection; and a heatmap-based digital image processing method for multi-class defect interpretation. YOLOv8n is used as a visual confirmation layer, and GPS coordinates are assigned to confirmed defects for maintenance mapping. The filter-based method achieved 84.21% YOLO-confirmed roughness accuracy, the Random Forest model achieved F1 scores of 78.4% for potholes and 83.9% for cracks, and the heatmap-based method achieved 90.3% overall YOLO-confirmed accuracy.</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

Investigation of road surface defects using robust techniques with multi-sensor integration

  • 期刊论文

  • 发表日期：

    2026年7月20日

查看全文
  • Muhammad Shahzad Alam Khan,
  • Hammad Majeed,
  • Anas bin Aqeel

摘要

Reliable detection of road surface defects is important for traffic safety, pavement maintenance, and cost-effective infrastructure management, especially where repeated manual inspection is expensive. This paper presents a low-cost multi-sensor framework for automated and geo-localized road assessment using 2D LiDAR, a monocular camera, GPS, and an inertial measurement unit. LiDAR scans are converted into road-surface profiles, non-road structures are removed using density-based clustering and principal component analysis, and the low-frequency road undulation is separated from the raw elevation profile to extract high-frequency roughness signatures. Three detection strategies are evaluated: a filter-based roughness detector using autocorrelation, linear predictive coding, prediction power error, and roughness index; a Random Forest classifier for pothole and crack detection; and a heatmap-based digital image processing method for multi-class defect interpretation. YOLOv8n is used as a visual confirmation layer, and GPS coordinates are assigned to confirmed defects for maintenance mapping. The filter-based method achieved 84.21% YOLO-confirmed roughness accuracy, the Random Forest model achieved F1 scores of 78.4% for potholes and 83.9% for cracks, and the heatmap-based method achieved 90.3% overall YOLO-confirmed accuracy.