| Summary |
The research targets of this project are three concrete bridges, i.e., beam bridge, slab bridge, and box girder bridge. The goal is to utilize unmanned aerial vehicles (UAVs) to acquire images of several bridge components, such as the main girder, diaphragm girder, pillar, pillar cap, slab, abutment, and wing wall. Later, AI machine learning semantic segmentation technology was applied to detect several bridge defects, such as concrete cracks, concrete spalling, rebar corrosion, infiltration, and efflorescence. In summary, the research topics and significant issues include the following parts: literature review, development of machine learning technology, development of UAV technology,
bridge three-dimensional (3D) modeling, and technology promotion.
In the development of machine learning technology, we have cleaned up the TBMS2 database and chose 4,262 photos for defects labeling and DRU editing and assurance, in which a total of 3,304 images have been revised, which occupied 77.48%, based on the consistency of D, R, and U values in the specification and suggestions from the experts. For accuracy evaluation purposes, 4,805 data points were collected from 14 bridges in the database. After the same data cleaning procedure, 833 data points remained. This project developed two deep learning semantic segmentation models, including DeepLab v3++ and Lawin, as well as a DRU estimation model. After accuracy assessment and testing, the developed segmentation methods have high feasibility in bridge defect detection.
However, though the developed DRU estimation model has high potential in real applications, its accuracy still needs to be improved because the DERU scoring standard is varied, and the number of training data and their diversity are
insufficient.
In the development of UAV technology, this project has developed a data fusion positional algorithm based on
VIO, IMU, RTK, and UWB that can conduct self-flying and image acquisition under a GNSS-denied environment. It
offers the UAV a global and consistent positional result. After several preliminary experiments, an absolute positional
error of less than 30 cm was obtained. In the meantime, this project has designed a Y6 UAV dedicated to bridging
inspection. It is also equipped with a gimbal capable of 180 degrees of up-down rotation angle and an LED light. We
use a bridge 3D model for flight path planning. |
