Robust Detection of Crack Waveform Using Images for Automatic Road Repair System

Abstract

Paved roads are said to be able to use them for a long time with adequate maintenances on a regular basis. However, due to a shortage of manpower for repairs, sufficient maintenance have not been taken on some municipal roads in Japan, and the deterioration has progressed rapidly. Furthermore, around 2025, many asphalt pavements paved during the period of high economic growth in Japan are expected to deteriorate rapidly, hence, a more effective method than the current manual repairing method is required. In this research, we aim to realize a machine that automatically detects and repairs cracks. As a starting point, this paper studies crack detection, one of the important elemental technologies. Assuming actual use, it is necessary to be robust to disturbances such as different colors of paved roads and illumination change due to the weather. To solve this problem, we propose a robust crack detection using deep learning. Specifically, the crack detection was performed by classifying cracked and uncracked areas by semantic segmentation using U-Net. The learning process was performed with various images including lighting changes in the training data set. As a result, we achieved a robust segmentation of cracked areas with 92.5 percent accuracy of Intersection over Union.