Binocular video-based 3D reconstruction and length quantification of cracks in concrete structures

Abstract

In the past few years, image-based methods for crack inspection have been developed to reduce the cost of the manual crack inspection. However, current image-based methods usually obtain only localized planar crack detection results, while global location and three-dimensional (3D) geometric information of the crack is desired to evaluate the whole structure. In addition, automatic quantification of the actual sizes of all cracks in the global structure is a challenge to overcome. To address these issues, this paper develops an automated 3D reconstruction and length quantification framework for cracks in concrete structures based on binocular videos. In the proposed framework, a crack semantic 3D reconstruction method, combined with binocular visual simultaneous localization and mapping (VSLAM) and a high-performance segmentation network, is first proposed to achieve accurate crack 3D characterization and global localization. A 3D crack quantification method based on point cloud processing is also developed to accurately identify individual cracks on the global structure and quantify their lengths. Field tests are conducted on a concrete bridge to demonstrate the automated inspection framework, which shows high efficiency and practicability. The accuracy of crack detection and quantification of the proposed method is also validated against the results of manual inspection.