Graph-based change detection of pavement cracks

Abstract

Pavement crack deterioration threatens road safety, but current maintenance strategies rely on composite indicators that lack crack location and attribute changes, failing to accurately track deterioration. Traditional feature-point-based methods struggle with temporal crack correspondence due to noise and shape variability. However, local structural features like intersections and inflection points are significant and extractable, providing a reliable basis for crack correspondence. This paper proposes an unsupervised graph-based framework for crack change detection. First, a curvature-first distance-optimized algorithm extracts stable keypoints to construct crack graphs, representing structural information. Second, a graph matching strategy combines Bezier curve similarity and Monte Carlo Tree Search to resolve structural correspondences, enabling accurate change detection. To address data scarcity, a Voronoi-based simulator models crack propagation through controlled stress fields. Experiments on synthetic and real-world datasets achieved crack change detection accuracies of 97.95% and 90.18%, respectively, demonstrating high accuracy without relying on learning-based components.