Abstract:
In South Africa it is a requirement for all road owners to conduct principle visual bridge inspections of all structures every five years, as per TMH 19. A principal inspection is a comprehensive visual inspection of the entire structure and is conducted by qualified bridge inspectors. This dissertation introduces the application of new Fourth Industrial Revolution (4IR) technology solutions into the realm of bridge inspection methodologies in South Africa, aimed to enhance the current bridge inspection methodology, while considering the cost and time components. For this study image data for eight bridge and culvert structures were captured using Unmanned Aerial Vehicles (UAVs). Point cloud models were created from the captured images by using photogrammetry software. Accredited bridge inspectors were approached to complete new inspection sheets of the bridge structures using only the point cloud models and captured images, as a proposed new inspection methodology. The cost and time components of the new inspection methodology were recorded and the cost and time components of traditional TMH 19 inspections were analysed. Existing inventory and inspection images of bridge roadway joints were compiled to develop different Convolutional Neural Network (CNN) models and the possibility of classifying bridge defects autonomously were considered. This dissertation compares historic inspection ratings to the new inspection ratings to investigate the effectiveness and practicality of the new proposed inspection methodology. The cost and time components of the new inspection methodology and traditional TMH 19 inspections were compared to determine if the new proposed inspection methodology prove to have any cost- and time benefits. The prediction results of the different CNN models were compared and analysed to determine if it is possible to detect and classify bridge defects autonomously, using existing image data and applying deep learning and computer vision techniques. The study concluded that bridge inspectors can use point cloud models and captured images to complete inspection sheets, as a proposed new inspection methodology. Bridge inspectors are able to identify and rate critical defects, but there are limitations to the application of the new methodology and specific use cases need to be identified. The time- and cost-saving aspects of the new inspection methodology did not prove to have any benefits and depend on limiting the human aspect of inspections, possible through computer vision and deep learning applications to identify defects autonomously. The best performing CNN model utilises transfer learning and data augmentation to predict with 95% accuracy from images if a bridge roadway joint has a defect and with 65% accuracy if the bridge roadway joint has no defect.