Automated crack segmentation and 3D dimension estimation using monocular RGB images for road infrastructure monitoring

Measurement, vol. 291, pp. 123084

Abstract

Accurate crack detection and dimensional quantification are critical for infrastructure maintenance, yet existing approaches often lack physical scale information or rely on complex sensing setups. This paper presents MonoCrackDepthNet (MCDN), a monocular RGB-based framework for crack segmentation, metric depth estimation, single-view 3D reconstruction, and scale-aware crack quantification. MCDN employs a Hierarchical Feature Encoder built upon a hierarchically structured vision transformer backbone, a Multi-Fusion Decoder, and a dual-output Mask-Depth Head to generate spatially aligned crack masks and canonical inverse-depth maps. Metric depth is learned through focal-length-normalized supervision that combines scale-and-shift invariant structural constraints with an unaligned metric loss, while geometry-consistent RGB-D pitch augmentation introduces controlled camera-angle variations during training. The canonical prediction is converted to native-camera metric Z-depth and back-projected using the documented camera intrinsics, enabling physical crack measurement without multi-view imagery, calibration objects in the measurement scene, or additional sensors. Experimental results show that MCDN achieves state-of-the-art-level segmentation accuracy (62.43% IoU and 76.88% F1-score) and a mean dimensional relative error of 8.75%, corresponding to a 48% reduction in measurement error compared with the binocular-camera and checkerboard-based methods. Field testing further yields relative errors of 6.4% for crack length and 9.7% for width, demonstrating the practical applicability of MCDN for automated pavement inspection.

Authors 4

1. Yuanyuan Hu

   Southeast University

   Affiliation as printed

   Intelligent Transportation System Research Center, Southeast University, 35 Jinxianghe Road, Xuanwu District, Nanjing, 210096, China

2. Hancheng Zhang Aachen Institute of Highway Engineering

   Wenzhou Business College · RWTH Aachen University

   Affiliation as printed

   Institute of Highway Engineering. RWTH Aachen University, Mies-van-der-Rohe-Straße 1, Aachen, 52074, Germany

   School of Information Engineering, Wenzhou Business College, Wenzhou, 325035, China

3. Jing Hu

   Southeast University

   Affiliation as printed

   Intelligent Transportation System Research Center, Southeast University, 35 Jinxianghe Road, Xuanwu District, Nanjing, 210096, China

4. Pengfei Liu corresponding Aachen Institute of Highway Engineering

   RWTH Aachen University

   Affiliation as printed

   Institute of Highway Engineering. RWTH Aachen University, Mies-van-der-Rohe-Straße 1, Aachen, 52074, Germany