Quantitative Road Crack Evaluation by a U-Net Architecture using Smartphone Images and Lidar Data

Abstract

<p> Road cracks are an important concern of administrators. Visual inspection is labor-intensive and subjective, while previous algorithms detecting cracks from optical camera images were not accurate. Furthermore, the actual length and thicknesses of a crack cannot be estimated only from images. Light Detection and Ranging (Lidar) is a standard feature introduced on the latest smartphones. In this research, for completely automatic, accurate and quantitative road crack evaluation using smartphones, an up-to-date segmentation technique, U-Net with morphology transform adopting data augmentation was proposed. Lidar 3D point cloud data of smartphones is linked to color data obtained from cameras. By registering images to Lidar data, geometrical relationships were estimated to calculate the length and thicknesses. The proposed algorithm was validated by a standard database of road cracks and dataset constructed by the authors, showing 95% length accuracy and 0.98 coefficient of determination for thickness estimation irrespective of various crack shapes and asphalt pavement color patterns. </p>