Answers are grounded in this paper and cite their sources.
Advanced Structural Health Monitoring (SHM) is fundamentally reliant on the autonomous assessment of structural integrity. Nonetheless, the precise segmentation of hairline cracks against severe background disturbance and achieving high-fidelity three-dimensional (3D) reconstruction under lightweight computational constraints, in complex civil engineering environments, remains challenging. To address these bottlenecks, this paper proposes a unified computer vision framework integrating a morphology-enhanced foundation model with a robust Gaussian Splatting Simultaneous Localization and Mapping (GS-SLAM) system. Firstly, a morphology-aware fine-tuning Segment Anything Model (SAM) is proposed to segment two-dimensional (2D) hairline cracks. By incorporating a morphology-sensitive weighting strategy, it overcomes extreme class imbalance to precisely extract continuous crack topologies from highly textured backgrounds. Subsequently, in the 3D domain, a Frequency-Aware GS-SLAM (FreqGS-SLAM) is proposed for high-fidelity structural scene reconstruction. By developing a Frequency-Aware Adaptive Photometric (FAAP) loss coupled with a semantically guided densification mechanism, the system reduces the sensitivity of Gaussian-map optimization to tracking-induced high-frequency photometric residuals, explicitly allocating Gaussian primitives to defect regions to prevent high-frequency detail degradation. The proposed framework has been comprehensively validated on public datasets, large-scale urban road structures, post-earthquake ground fissures, and a controlled full-scale exterior-wall crack. Experimental results demonstrate improved crack-detail preservation and a favorable reconstruction-efficiency trade-off under the evaluated workstation configuration. In one controlled close-range exterior-wall case, the reconstructed crack width at a predefined locally wide and caliper-accessible position was 2.41 mm, compared with a digital-caliper reference of 2.53 mm, giving a case-specific absolute difference of 0.12 mm. This single-location result is not interpreted as the general geometric resolution or field measurement accuracy of the proposed framework. The present implementation supports sequential online pose estimation and incremental map construction followed by post-mapping refinement, while real-time deployment on UAV or mobile edge hardware remains a subject of future investigation.
No takes yet. Share an insight, caveat, or question.
Xu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: