基于Boosting-Monodepth的管道病害深度估计与三维重建

Pipeline Damage Depth Estimation and 3D Reconstruction Based on Boosting-Monodepth Algorithm

摘要:

城市地下管道是城市的血脉经络，但随着排水管道的大量投入运营和使用年限增加，引发了一系列的管道病害安全隐患，如管道整体结构变形、内表面破裂和管中异物插入等问题，传统的病害图像视频采集、检测和后期病害分类甄选都是从二维视角出发，欠缺对三维空间信息（深度）的考虑。针对上述3种病害从生成深度图、由二维深度图重建三维管道病害这两方面进行研究，提出了一种基于boosting-monodepth的双重深度估计方法以提升深度图效果，最终生成画面连续一致、轮廓清晰的深度图。性能评估方面采用Abs-Rel、RMSE、SqRel、ORD和D3R等通用指标，与传统算法对比，结果显示boosting-monodepth的RMSE值降低了30%，精确度指标δ＜1.25时，模型深度信息预测精确度提高了18%，此后以得到的深度图为基础重建管道病害三维点云，并在CloudCompare软件上三维可视化，最后采用随机采样一致算法测算病害深度并和实测数据对比证明其有效性和准确性。

Abstract:

Urban underground pipelines are the blood channels of the city. However， with the large number of drainage pipelines put into operation and their service life increasing， a series of hidden dangers of pipeline diseases have arisen， such as the overall structural deformation of the pipeline， internal surface cracking， and the insertion of foreign matters in the pipeline. Traditional disease image video capture， detection， and later disease classification and selection are conducted from a two-dimensional（2D） perspective， lacking consideration of three-dimensional spatial information （depth）. Aimed at the above three diseases， an experiment is conducted from generating depth maps and reconstructing three-dimensional（3D） pipeline diseases from 2D depth maps. A dual depth estimation method based on boosting-monodepth is proposed to improve the effect of depth maps， and finally a depth map with continuous and consistent images and clear outlines is generated. In terms of performance evaluation， Abs-Rel， RMSE， SqRel， ORD， D3R and other general indexes are compared with those of the traditional algorithm， and the results show that the boosting-monodepth is reduced by 30% at RMSE and increased by 18% at a threshold of δ< 1.25. Afterwards， the 3D point cloud of pipeline disease is reconstructed based on the depth map obtained， and 3D visualization is performed on CloudCompare. Finally， the disease depth is calculated using the random sampling consistent algorithm and compared with the real measured data to prove its effectiveness and accuracy.