Geometry-Consistent Alignment for Time-Lapse Crack Monitoring in Field-Based SHM

Keypoint-based detectors such as SIFT and SURF—built on Gaussian scale-space—are widely used in image-based structural health monitoring (SHM), but frequently fail under field conditions. Their reliance on isotropic smoothing suppresses fine crack edges and yields unstable alignment under occlusion, surface texture, or varying illumination. Binary descriptors like ORB and BRISK, while faster, inherit similar drawbacks due to Gaussian pyramids and weak structural saliency. This study presents a geometry-consistent alignment framework that addresses these challenges by repurposing anisotropic diffusion–based feature detection and integrating it with RANSAC-driven homography estimation. Compared to the isotropic blurring of Gaussian filtering, anisotropic diffusion constructs a nonlinear scale space that selectively smooths homogeneous regions while preserving crack-relevant discontinuities—enhancing keypoint stability under low contrast and visual clutter. RANSAC further performs robust outlier rejection to fit a reliable geometric transformation under sparse or noisy conditions. The proposed pipeline is unsupervised, training-free, and calibration-free—designed for crack imagery acquired by UAV or handheld platforms. Validated on over 100 field-acquired image pairs across varied degradation scenarios—including textured masonry, strong perspective distortion, and moving shadows—the method achieved average area errors below 5% and spine length errors under 15% relative to manual ground truth. In comparison, classical methods frequently exhibited 2–4× higher errors; for instance, SURF showed over 25% length error in occluded scenes due to clustered inliers, while SIFT failed to recover alignment under similar conditions. Fully interpretable and lightweight, the framework supports scalable crack evolution tracking without data annotation or parameter tuning. Its compatibility with mobile sensing and transparency in edge preservation lay the foundation for integration into smart SHM workflows and adaptive control systems. Future directions include stereo or depth-based correction for non-planar surfaces and integration with learned descriptors—extending the interpretable, calibration-free foundation established in this work. Furthermore, unsupervised mapping of crack evolution patterns from aligned imagery may enable cluster-based or topology-informed monitoring strategies.