Computer Vision Applications in Structural Engineering

This thesis explores the applications of computer vision technologies in structural engineering, focussing on four key studies as follows. 1. Deep Learning Based Fire Risk Detection on Construction Sites: The recent large-scale fire incidents on construction sites in South Korea have highlighted the need for computer vision technology to detect fire risks before an actual occurrence of fire. This study developed a proactive fire risk detection system by detecting the coexistence of an ignition source (sparks) and a combustible material (urethane foam or Styrofoam) using object detection on images from a surveillance camera. Statistical analysis was carried out on fire incidences on construction sites in South Korea to provide insight into the cause of the large-scale fire incidents. Labeling approaches were discussed to improve the performance of the object detectors for sparks and urethane foams. Detecting ignition sources and combustible materials at a distance was discussed in order to improve the performance for long-distance objects. Two candidate deep learning models, Yolov5 and EfficientDet, were compared in their performance. It was found that Yolov5 showed slightly higher mAP performances: Yolov5 models showed mAPs from 87% to 90% and EfficientDet models showed mAPs from 82% to 87%, depending on the complexity of the model. However, Yolov5 showed distinctive advantages over EfficientDet in terms of easiness and speed of learning. 2. Structural Shape Monitoring Using an RTK enabled UAV with GCP-Free Georeferencing: This study investigates the feasibility of the structural shape monitoring approach consisting of a Real-time Kinematic (RTK) enabled Unmanned Aerial Vehicle (UAV) and georeferencing without Ground Control Points (GCPs), expected to achieve a high efficient in field work. Quantitative assessments were carried out on 1) the positioning accuracy of the georeferenced point clouds generated by the shape monitoring system without GCPs and 2) the shape change detection capability of the shape monitoring system. The positioning accuracy of the point clouds generated by the system was evaluated by an RTK rover. The disagreements between them were found to be less than 2 cm. Shape change detection was carried out for detecting a brick shape of different depths of 8mm, 20mm, 44mm, and 84mm, attached to a wall of a target structure. The brick shapes of 20mm depth or higher were detected successfully, but that of 8mm depth was not, showing the detection limit of the current system configuration. The depth and volume estimation were carried out and the maximum estimation errors of them were found to be 11%. The registrations on the point clouds were carried out to improve the alignments between the point clouds to obtain better depth and volume estimations as well as observing the georeferencing accuracy. After registration, the maximum error was reduced from 11% to 6%. The above results suggested that the structural shape monitoring approach has the potential to improve the practice of visual inspection of infrastructure by reconstructed point clouds accurately and efficiently by an RTK enabled UAV and GCP-free georeferencing. 3. RTK-UAV based GCP-free Structural Shape Monitoring on Tall Industrial Chimneys: This study investigated the feasibility of structural shape monitoring in tall industrial chimneys using a Real-Time Kinetic (RTK) enabled Unmanned Aerial Vehicle (UAV). The advantage of using an RTK-enabled UAV is the extreme efficiency in field survey work, mainly due to the georeferencing and autopilot feature based on the RTK. This automatic georeferencing also makes any further processing and quantification on the point clouds straightforward, as the dimension in the point clouds is the same as in the real structure of interest. Experiments on a scaled chimney model were carried out with artificially induced tilts and settlements. Eight autopilot surveys were carried out to produce the georeferenced point clouds, and the horizontal and vertical displacements of the top of the chimney were estimated with and without artificial markers. In order to estimate displacements between two point clouds without artificial markers, the point cloud registration approach was used, with the rigid-body motion assumption that the part of the point clouds to be monitored undergoes a rigid body motion only, without any deformation. Estimated horizontal and vertical displacements were found to agree well with the actual displacements and the maximum estimation errors were found to be around 1cm for both cases, with and without the artificial markers. This result suggested that the registration approach can be used without any artificial markers on tall industrial chimneys. 4. RTK-UAV based 3D Crack Identification, Visualisation and Quantification for Structural Health Monitoring: This study proposed enhancements for 3D Crack Identification, Visualisation, and Quantification in Structural Health Monitoring using a Real-Time-Kinematic (RTK) enabled Unmanned Aerial Vehicle (UAV). The proposed enhancements are for the 3D Crack Identification approach that combines 2D crack segmentation results with reprojection, available during 3D reconstruction. In the crack identification approach, each 3D point in a point cloud of a structure is reprojected onto the image coordinates, where the 3D point is classified as either a crack point or non-crack point based on the 2D crack segmentation result. In order to minimise false-positive crack identification in the approach, it is proposed to use the hidden point removal after depth filtering. In addition, the consistency index was proposed to handle inconsistent results of 2D segmentation on multiple photos for the same crack on the structure. The use of an RTK-enabled UAV is proposed to make crack width estimation extremely efficient without ground control points (GCPs) or scale-bars in photos. The proposed enhancements were validated on a field experiment in a real building with an artificial crack. The results showed that the proposed combined method was very efficient in minimising false-positive crack identification. The 3D crack identification results were found to be highly influenced by the consistency index. The consistency index of 2 was found to have the lowest error for the experiment. The point cloud reconstructed with GCP-free georeferencing made crack width estimation straightforward in the 3D point cloud.<p></p>

Author dc:creator

• H Ann (22000658)

• Computer vision
• Deep learning
• RTK UAV
• Structural health monitoring
• Civil engineering

Statement dc:rights

• All rights reserved
• Open Access after 2028-01-27

Identifier

10871/139753

OAI identifier oai:identifier

oai:figshare.com:article/29843807