Scholarlink

<p>Existing pipeline inspection methods fail to provide quantitative measurements of critical three-dimensional (3D) geometric properties (e.g., defect depth, volume) despite the high accuracy. To address this limitation, a 3D reconstruction method for sewer pipelines based on monocular vision is proposed to generate depth maps and 3D models of pipeline defects from standard images. First, a dual-boost depth estimation approach is developed to enhance the quality of depth maps, ultimately producing continuous, consistent depth maps with clear object contours. Experimental results demonstrate that the measurement error (RMSE) of the proposed method is reduced by 30% and the accuracy metric δ (with a threshold of δ &lt; 1.25) is improved by 18% compared to conventional methods. Second, depth information of the generated depth maps is used to reconstruct 3D point clouds of pipeline defects. Finally, a combination method of voxel downsampling and statistical filtering is employed to eliminate noise and obvious outliers from the point clouds. The comparison results show that the relative errors between the reconstructed dimensions and the physical measurements are all below 3.5%, demonstrating the high accuracy and metrological validity of the proposed measurement system.</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

Monocular depth estimation and 3D reconstruction for defects in sewer pipelines

  • 期刊论文

  • 发表日期：

    2026年8月4日

查看全文
  • Hao Hu,
  • Duo Ma,
  • Niannian Wang,
  • Haobang Hu,
  • Kangjian Yang,
  • Xue Jiang

摘要

Existing pipeline inspection methods fail to provide quantitative measurements of critical three-dimensional (3D) geometric properties (e.g., defect depth, volume) despite the high accuracy. To address this limitation, a 3D reconstruction method for sewer pipelines based on monocular vision is proposed to generate depth maps and 3D models of pipeline defects from standard images. First, a dual-boost depth estimation approach is developed to enhance the quality of depth maps, ultimately producing continuous, consistent depth maps with clear object contours. Experimental results demonstrate that the measurement error (RMSE) of the proposed method is reduced by 30% and the accuracy metric δ (with a threshold of δ < 1.25) is improved by 18% compared to conventional methods. Second, depth information of the generated depth maps is used to reconstruct 3D point clouds of pipeline defects. Finally, a combination method of voxel downsampling and statistical filtering is employed to eliminate noise and obvious outliers from the point clouds. The comparison results show that the relative errors between the reconstructed dimensions and the physical measurements are all below 3.5%, demonstrating the high accuracy and metrological validity of the proposed measurement system.