A hybrid method for pavement crack width measurement

Abstract

Accurate crack width measurement is crucial to determine the severity level of pavement cracks and the selection of crack repair strategy. We propose to combine the shortest method with the orthogonal projection method to produce a novel hybrid method. The hybrid method obtains the crack width by identifying a pair of points that give the shortest distance while being close to the orthogonal direction. We tested quantitatively and qualitatively the shortest, orthogonal projection and the hybrid methods on images taken from our CrackSkel700 dataset and two other open datasets, CrackForest, and Crack500. For the qualitative test, one hundred manual measurements are taken randomly from five images in the datasets to compare the shortest, orthogonal projection and hybrid methods on real cracks. Compared to the shortest and orthogonal projection methods, the hybrid method obtains the least root mean squared (RMS) error of 1.000 pixels and the least absolute pixel deviation of 0.732 pixels. Then, we generated 30 synthetic cracks using circles with a known diameter ranging from 3 pixels to 61 pixels to evaluate the accuracy of each method. The synthetic cracks generated a total of 18,404 ground truth measurements. Based on the synthetic cracks, the hybrid method obtains the least average absolute deviation of 1.769 pixels and the highest correlation coefficient of 0.956 compared to the shortest and orthogonal projection method on synthetic cracks. The qualitative comparisons of real and synthetic cracks show that the hybrid method improves the shortest method significantly by reducing the number of repeated measurements. On the other hand, the hybrid method improves the orthogonal projection method by reducing the overestimation of non-parallel and high curvature cracks. Hence, we show that the hybrid method generalizes better to more crack patterns and thus produces more accurate crack width estimation than the orthogonal projection and the shortest method.