Pavement crack detection using non‐local theory and iterative sampling

Abstract

Crack is a common form of road distress and a key study of an intelligent transportation system. However, automatic pavement crack detection is a very challenging task due to noisy texture background, intensity inhomogeneity, and topology complexity. In this paper, a new pavement crack detection algorithm to address these issues is proposed. First, non-local block matching strategy and local statistical mean are put together to generate the probability map of cracks, which has advantages on automatic threshold choosing and strong resistance to intensity inhomogeneity. Second, an iterative seed points sampling algorithm is proposed, which makes full use of the area and shape of connected regions where the seeds lie in, thus exploiting high reliable crack seeds for following curves extraction. Finally, a minimum spanning tree (MST) is adopted to connect points into crack curves and employ a crack growth method to find out the cracks, which is specified to deal with complex topology of cracks. For parameters, a robust and optimal parameters selection rule is obtained by data driven method. The algorithm is compared with other state-of-the-art algorithms on two datasets. The experiment result shows that the proposed method has a better detection performance on $F1$-measure score over other methods.