VISION-BASED APPROACHES for QUANTIFYING CRACKS in CONCRETE STRUCTURES

Abstract

In this paper, a combination of photogrammetric, computer-vision, and deep-learning approaches are proposed for accurate detection and quantification of cracks from the images of concrete structures. In particular, a semantic segmentation approach using UNet is applied, which is trained on a customized dataset of real-world images. Then, two photogrammetric methods are assessed for reconstructing the full figure of the cracks from stereo images. One approach is based on detecting the dominant structural plane surrounding the crack and projecting the crack pixels to this 3D plane. The second approach is based on matching the crack pixels across two images. To be able to perform the 3D reconstructions accurately, a rigorous calibration of the intrinsic calibration parameters of the cameras is performed. The relative orientation parameters between the stereo cameras are also determined in the calibration procedure. Extensive experiments are performed to evaluate each phase of this detection-and-quantification workflow. In general, cracks can be detected with an average precision of 87.48% and recall of 87.45%. They can be reconstructed in 3D with an accuracy as high as 0.05 mm.