Study on a vehicular defect identification system for girder bottom inspection of bridges

Abstract

The girder bottom inspection is becoming an important part of the bridge maintenance process. In this study, a vehicular defect identification system was built to make the inspection process of bridge bottoms more intelligent, efficient and accurate. The system contains three main parts: image acquisition, image stitching and defect recognition. The image acquisition part was responsible for controlling the start and stop of image acquisition, data transmission and image storage. The image sequences collected were processed and stitched into a panoramic image during the image stitching process, and the coordinate systems of images would also be unified. Finally, the defects in the image were recognized and positioned. Combined with the BIM model, multiscale digital display of bridge bottom defect, including defect recognition and positioning results, was obtained. With the multiscale information, the maintenance for bridges will become more convenient. The deep learning model U2-Net was used to detect cracks and realized a defect detection accuracy of millimeter-level. The experimental results proved that the cracks in the images of the bridge bottom could be detected effectively using the proposed method with a high performance of 79.15 % test dataset F1-score and 0.691 MIoU. Additionally, the proposed defect location method had a centimeter-level defect location accuracy.