In order to join virtual sessions, you must be registered and logged-in(Were you registered for the in-person meeting in Denver? if yes, just log in.) 
Note: All session times are in Mountain Daylight Time.

Deep Learning-based Semantic Segmentation of 3D Point Clouds: A Case Study for Hydraulic Structures

Authors: Tianyang Chen*, Center for Applied Geographic Information Science, the University of North Carolina at Charlotte, Wenwu Tang, Center for Applied Geographic Information Science, the University of North Carolina at Charlotte, Shen-En Chen, Dept. of Civil and Environmental Engineering, the University of North Carolina at Charlotte, Craig Allan, Center for Applied Geographic Information Science, the University of North Carolina at Charlotte, John Diemer, Dept. of Geography and Earth Sciences, the University of North Carolina at Charlotte, Zachery Slocum, Center for Applied Geographic Information Science, the University of North Carolina at Charlotte, Tarini Shukla, Center for Applied Geographic Information Science, the University of North Carolina at Charlotte, Navanit Sri Shanmugam, Department of Civil and Environmental Engineering, the University of North Carolina at Charlotte, Vidya Subhash Chavan, Department of Civil and Environmental Engineering, the University of North Carolina at Charlotte, Matthew S. Lauffer, Hydraulics Unit, North Carolina Department of Transportation
Topics: Geographic Information Science and Systems, Water Resources and Hydrology, UAS / UAV
Keywords: 3D point cloud, semantic segmentation, outdoor scenes,deep learning
Session Type: Paper
Presentation File: No File Uploaded


Semantic segmentation of 3D point clouds can be applied in different domains with an increase in LiDAR data. Deep neural networks directly using point cloud as inputs have been developed to handle such an application. This deep neural network technology often outperforms other Convolutional Neural Network-based methods, which requires to transform point clouds to regular 3D voxel grids. Researchers have developed deep neural networks upon such an approach for their needs of semantic segmentation in different contexts, such as road marking detection, animal detection, and vehicle detection. In this study, we utilize this deep learning approach in the context of hydraulic structure detection. Specifically, we aim to automatically detect hydraulic structures (e.g. bridge, pipe, culvert, etc.) and their components (e.g. piers, guardrails, retaining walls, etc.). However, rather than a general network development relying on benchmark datasets, the development for a specific context, e.g., hydraulic structures in our case, can face a challenge in terms of the availability of training data. We leverage transfer learning by adopting the related benchmark datasets to resolve this challenge. Moreover, we use our training data to further train the deep neural network, which is collected by the LiDAR instrument from fieldwork at Charlotte, NC and is labeled manually. This case study can benefit the use of deep learning-based semantic segmentation in the context of hydraulic structures and provide empirical knowledge to studies in related contexts.

To access contact information login