Bridge Scour Detection Using Terrestrial LiDAR and Advanced Quantification Techniques
Analytics
17 views ◎9 downloads ⇓
Abstract
Scour is an important factor affecting the hydraulic structures of a bridge. Remote sensing techniques such as terrestrial LiDAR (Light Detection And Ranging) can help speed up the inspection process and provide high-resolution records of the extent of scour. With LiDAR point cloud data, a temporal record of scour can be established. However, there are limitations to LiDAR scans. For example, a scan would contain not just the scour but surfaces surrounding the scour as well. Thus, there is a need to identify and separate scour points from the rest. Moreover, laser light does not bend and can be obstructed by objects along the light path resulting in missing geometric information behind the obstacles thereby creating a void in the point cloud data. To address this data void issue and to ‘reconstruct’ likely scour void, innovative analytical processes are being explored in this dissertation. 1. To automate scour detection and classification, 3D Point Capsule Network (3D PCN) for processing LiDAR point clouds captured from bridge hydraulic structure scans is presented. Scan results were first processed to cut portions that contain scour points. Synthetic data resembling a scour were then generated and 3D PCN, powered by a dynamic routing algorithm, was used to label the points of a given scour point cloud into scour and non-scour points. If scour is identified, it is segmented (cut) out from the point cloud for documentation. 2. To ‘fill in’ the missing data, spatial interpolation of 3D LiDAR point cloud data using Ordinary Kriging (OK) method is suggested and actual field data from scanning a scoured bridge pier is presented to demonstrate the application. Kriging is a geostatistical interpolation technique and OK assumes that the spatial variation of the phenomenon or object being considered is random and intrinsically stationary with a constant mean. Here, the complete scour envelope is reconstructed using kriging. 3. Interpolation of the point cloud data can result in either extremes of data density, very dense or very sparse. A method to find an ‘optimum’ point-to-point distance after interpolation using processing times, surface area and volume calculations is presented. Scanned point cloud from the Phillips Road Bridge of the Toby Creek, Charlotte, North Carolina, has been used for the study. The different processes (OK and 3D PCN) are then applied to the point cloud data set separately. The results from the distinct methodologies are summarized as follows: • The 3D PCN was trained to detect scour using 1,000 sets of synthetic scour data, using a split of 750-150-100 for training, evaluation, and testing. • The resulting model had an accuracy of 63% in identifying scour points from the input scour and non-scour point cloud. • The network performed well on a real-world point cloud from the Phillips Road Bridge pier scour. • Data voids were identified on the same real-world scour and OK process was used to fill in those voids. • Post-kriging spatial resolution of the points was much higher (i.e., point-to-point distances were much lower) than the original, which had varying point density in different portions of the cloud. • Scour depth was measured, and surface areas and volumes were calculated for scenarios consisting of nine different spatial resolutions. A point-to-point distance of 20mm was found to be the optimal spatial resolution considering total processing times and comparison of the scour parameters with the actual values.