Image registration for bridge defect growth tracking

Worldwide infrastructure assets are ageing, and the industry requires reliable condition information. Visual bridge inspections include digital photographs, typically taken from different angles. Without pixel-wise correspondence between images, it is not possible to quantify and compare defect extent in an automated and objective manner. We propose a methodology to address this, and hence use historical defect images to their full potential to track defect growth. Given two photographs of a particular defect, we use a computer vision image registration algorithm to identify matching keypoints across the pair of images. A linear transformation is computed from the matched keypoints and applied to one of the images in order to obtain pixel-wise correspondence, whence the two images may be overlaid and compared. We apply absolute difference between RGB values to highlight the visual changes and hence direct the user’s attention towards potential alterations in condition. As the number of defects to monitor is increasing rapidly, this tool aims to alleviate the mounting pressure on bridge inspectors, who select which defects should be brought to an engineer’s attention. We identify the limitations of relying on single photographs as means of visual defect data recording. In some cases there are not enough keypoint matches, for example due to uniform surfaces, extreme changes in lighting conditions, or changes to surface colour (wet areas, paint or graffiti). For future inspections a more robust methodology is desirable, and we propose point cloud registration for this purpose.