3D modeling and automated measurement of concrete cracks via segment anything refinement and visual inertial LiDAR fusion

In practical applications, AI-based concrete crack inspection still suffers from performance degradation in few-shot, unfamiliar scenarios and lacks the capability for high-precision, synchronized quantification of three-dimensional (3D) crack geometry and location without manual post-processing. To address these limitations, a systematic methodology for crack segmentation, 3D reconstruction, and automated measurement is proposed, grounded in computer vision and Simultaneous Localization and Mapping (SLAM) techniques. First, a novel prompt generation strategy and a tailored segmentation quality assessment module are developed to improve the performance of the Segment Anything Model (SAM), enabling few-shot crack segmentation with strong generalization across diverse and unseen scenarios. Second, a comprehensive concrete cracks reconstruction within a 3D representation is achieved through a newly proposed Visual Inertial LiDAR (VIL) SLAM-based fusion approach. By integrating multi-frame RGB images, LiDAR point clouds, and inertial measurements, the method enables precise alignment of crack segmentation masks with 3D structural geometry, generating high-precise, dense, and semantically enriched point clouds that capture fine-grained crack details at real-world scale. Furthermore, an automated measurement module is introduced to directly quantify detailed crack geometrical and spatial information from the established 3D representation, eliminating manual post-processing and advancing beyond traditional image-based methods. Finally, extensive experiments are successfully conducted on diverse concrete structures validating the accuracy, robustness, and effectiveness in complex, non-planar, and cluttered environments of the proposed method.