ABSTRACT
Traditional methods for concrete bridge inspection are highly subjective, time-consuming, and in cases, not accurate. However, the advances in deep learning and convolutional neural networks (CNN) have revolutionized bridge inspection. Many studies have used CNNs to localize and quantify defects on the surface of structures which includes detecting delamination and rusting or measuring cracks and spalls. However, in most cases, the AI-based bridge inspection is designed for automated systems such as UAEs and robots. While these methods address many important issues in bridge inspection, such as accessibility and decreased labor, they also have shortcomings. The automated systems solely work based on the AI’s performance. In many cases, inspectors have to visit the site more than once to correct the mistakes of the AI. Moreover, the lack of data on concrete defects directly affects the accuracy of the CNN models and increases the post-processing time. This study proposes a methodology that allows the inspector to communicate with the AI on-site and in real-time. The method is designed to work on Mixed-Reality platforms to assist the inspector in localizing and quantifying concrete defects. Two different CNN models were selected and trained out of many different CNN architectures based on their accuracy, memory footprint, and performance speed. A Mixed-Reality user platform and tracking are also designed to allow the inspector to communicate with the AI at each step and also accurately save the location of the analyzed defects. The study also proposes a semi-supervised approach for increasing more labeled data and improving the accuracy of the defect quantification model. Lastly, a methodology for condition assessment of concrete defects using the Mixed Reality system is discussed. The inspection method in this study eliminates the subjectivity of inspection and guarantees human-verified results without the need for post-processing.
