Monitoring the conditions of hydraulic structures such as bridges and culverts is essential in warranting the safety and sustainability of transportation infrastructure. This is particularly important for North Carolina as more than 8 percent of NC bridges have been found in poor conditions and need immediate maintenance. Lidar and sonar technologies have been increasingly applied to support this monitoring need. However, the processing and classification of point cloud data generated from LiDAR and sonar techniques represents a challenge as hydraulic structures are often complicated in geometric characteristics and considerable labor and time are needed for the processing and classification.
To address this challenge, in this project, we developed DeepHyd, a deep learning-based 3D modeling framework and software tools for the automated classification of point cloud data of hydraulic structures. We collected field data from 11 sites in the Greater Charlotte Metropolitan region for training and validation of the deep learning algorithms. The field data collection combines the use of terrestrial LiDAR, sonar, total station, survey-grade GPS, and drone. The deep learning algorithm that we used for point cloud classification is a state-of-the-art 3D artificial intelligence technique. We used a two-tiered modeling approach to train deep learning algorithms using annotated point cloud data: classification of bridges from vegetation and ground, and classification of specific bridge components including beam, pier, railing, and retaining walls. We implemented scientific workflows to automate the classification of point cloud data of hydraulic structures using deep learning. Our major findings are: 1) our 3D deep learning algorithms in DeepHyd achieve high classification performance on point cloud data of hydraulic structures. 2) deep learning can effectively handle the classification of large volumes of point cloud data, but the training of deep learning algorithms requires large amounts of annotated data. 3) annotated point cloud data serve as a foundation database for the automated classification of hydraulic structures using artificial intelligence techniques. More annotated point cloud data, which cover alternative types of hydraulic structures, are needed for further improving classification performance.
