ABSTRACT
Bridge integrity is one of the most critical issues for infrastructure. Natural disasters or aging problems can cause damage to the structure and even affect safety, while regular inspection is one practical approach to affirm the current conditions of bridges. Because sophisticated inspectors are short with respect to the number of bridges, traditional visual inspection should be improved to be efficient and address the drawbacks of being time-consuming and labor-intensive. Additionally, specific bridge components have accessibility problems for inspection, e.g., abutments and bottom of bridge decks. Unmanned aerial vehicles (UAVs) have significantly enhanced the efficiency of bridge inspection in recent years. Therefore, this study proposes a systematic framework for bridge visual inspection utilizing UAVs in conjunction with deep learning and computer vision. This framework involves image localization, defect recognition, and damage quantification as three sequential tasks in bridge visual inspection. By shooting azimuths and camera-to-objective depths, bridge surface images captured from UAVs can be localized from the camera coordinates to the global coordinates (e.g., aligned with the GPS coordinates). Then, the defects on a bridge are detected by a deep learning model featuring instance segmentation, which enables cracks, spalling concrete areas, and exposed reinforcements to be highlighted in images. These images are also tagged with a point of interest and linked to the global coordinates. Moreover, the geometric parameters of these defects can also be interpreted through photogrammetry based on available azimuth and depth, and these defects are consequently quantified. An example using the proposed framework for bridge visual inspection is provided in this study for demonstration. As shown in the results, this framework further improves the efficiency of bridge inspection with added information, such as the three-dimensional locations, types, and geometric quantities of defects.
