A Matching Method for Longitudinal Cracks Based on Curvature Similarity

Pavement cracks captured byon-board cameras are distributed randomly in shapes, and only a part of the longitudinal cracks on the roads can be captured each timedue to the limited field of view, resulting in incomplete detection of longitudinal cracks. The imagesacquired by the on-board cameras are transformed from oblique images intoorthographic images by using the inverse perspective transformation method, thus the perspective distortionof the longitudinal cracks are corrected. Then a deep learning based semantic segmentation network, Deeplab V3+, is used to extract the pixels of cracks. Based on curvature similarity, a two-stage method from coarse to fineis proposed for matching longitudinal cracks.The crack curve to be matched is divided into a sequence of overlapping sub-curves, which are characterized by descriptor of curvature, and the matched sub-curves are the matched parts of cracks. The curvature is used to express the local shape and trend features of sub-curvesas descriptors, then the Kd-tree nearest neighbor matching algorithm is used to perform coarse and fast matching of thedescriptors. According to the spatial distribution of longitudinal cracks in two consecutive road images, constraints are added when the crack curves are divided into sub-curves, the starting point of the crack curve in previousimage and the ending point in the next image areused asterminus of each respective sub-curve. Based on the results of coarse matching, the interval of segmentation curves is gradually reduced, and the normalized cross-correlation coefficient is iteratively improved until it is greater than or equal to the threshold or the number of iterations exceeds the maximum value to achieve fine adjustment of the results of coarse matching. To verify the accuracy of the algorithm, a case study is carried out with different types of continuous and longitudinal cracks on the campus roads of Wuhan University.The minimum error of the matching results can reach 0.688 pixels. Compared with the coarse matching, the error after fine adjustmentreduces by 24.19% on average. In order to further verify the stability of the algorithm under noise, crack pixel noise is added to the simulation environment.When the standard deviation of Gaussian noise increases from 0 to 2 pixels, the error of the matching results increases by only 1.083 pixels. Compared with the SIFT algorithm, the proposed method can achieve successful matching in all ten groups of experiments, while the matching results of the SIFT algorithm completely fails in two groups. It indicates that the algorithm proposed has better stability under normal and noise environment.