A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
UnrollingNet: An attention-based deep learning approach for the segmentation of large-scale point clouds of tunnels
Abstract A novel projection-based learning method named UnrollingNet is developed to conduct a multi-label segmentation of various objects including seepage from 3D point clouds of tunnels. 3D laser scanning is first utilized to collect raw point clouds from the operating tunnels. An unrolling projection approach is created on the trimmed dataset to generate 2D representations. A U-net-based segmentation algorithm is employed to classify the tunnel at the pixel level. A pixel-weight cross-entropy loss together with a dual attention module is proposed to address the class imbalance issues and improve segmentation performance. A real cross-river tunnel section in China is used as a case study for demonstration. Results indicate that (1) the established model displays a high performance of point cloud projection, and the Purity score and Yield rate achieve 0.910 and 0.821, respectively; (2) the segmentation model performs well in multiple classes, the Intersection over Union (IOU), Precision, Recall, and F1 score for seepage segmentation achieve 0.66, 0.736, 0.864, and 0.795, respectively; (3) the model achieves better segmentation scores than other deep-learning-based point cloud segmentation models.
Highlights A deep learning method named TunnelNet is developed for 3D point cloud segmentation. A circle projection algorithm is proposed to transfer 3D point clouds into 2D images. An image segmentation model is presented via Unet coupled with DAN and novel loss. The developed method performs superiorly with high accuracy and great efficiency.
UnrollingNet: An attention-based deep learning approach for the segmentation of large-scale point clouds of tunnels
Abstract A novel projection-based learning method named UnrollingNet is developed to conduct a multi-label segmentation of various objects including seepage from 3D point clouds of tunnels. 3D laser scanning is first utilized to collect raw point clouds from the operating tunnels. An unrolling projection approach is created on the trimmed dataset to generate 2D representations. A U-net-based segmentation algorithm is employed to classify the tunnel at the pixel level. A pixel-weight cross-entropy loss together with a dual attention module is proposed to address the class imbalance issues and improve segmentation performance. A real cross-river tunnel section in China is used as a case study for demonstration. Results indicate that (1) the established model displays a high performance of point cloud projection, and the Purity score and Yield rate achieve 0.910 and 0.821, respectively; (2) the segmentation model performs well in multiple classes, the Intersection over Union (IOU), Precision, Recall, and F1 score for seepage segmentation achieve 0.66, 0.736, 0.864, and 0.795, respectively; (3) the model achieves better segmentation scores than other deep-learning-based point cloud segmentation models.
Highlights A deep learning method named TunnelNet is developed for 3D point cloud segmentation. A circle projection algorithm is proposed to transfer 3D point clouds into 2D images. An image segmentation model is presented via Unet coupled with DAN and novel loss. The developed method performs superiorly with high accuracy and great efficiency.
UnrollingNet: An attention-based deep learning approach for the segmentation of large-scale point clouds of tunnels
Zhang, Zhaoxiang (author) / Ji, Ankang (author) / Wang, Kunyu (author) / Zhang, Limao (author)
2022-06-22
Article (Journal)
Electronic Resource
English
Facet Segmentation-Based Line Segment Extraction for Large-Scale Point Clouds
| Online Contents | 2017