Automated Crack Detection Method Based on Deep Learning and 3D Reconstruction for Concrete Bridges

Automated image-based bridge crack detection, as a promising technique, can be used to overcome the limitations of human visual inspection. However, results from current image-based methods are generally localized and lack 3D geometric information, which makes it difficult for structural assessment. To solve this issue, a crack detection method that combines deep learning and 3D reconstruction is proposed in this paper. Firstly, a 2D feature-based approach is developed to extract keyframes from the video adaptively. Secondly, a segmentation network is implemented to conduct pixel-level crack segmentation. Finally, image-based 3D reconstruction and crack mapping are used to create the 3D structure model with crack semantics. A field experiment is also carried out on an in-service concrete bridge for validation and discussion of the proposed method. The 3D model created by the proposed method can significantly improve the crack inspection of concrete bridges.