ABSTRACT
Clay with different moisture contents has very different mechanical properties, which are vital for the stability of geotechnical structures. When desiccation cracks start to develop, the failure of soil structure might happen. Traditional image processing methods have been extensively applied to analyze the development of desiccation cracks in clay. However, these image analysis methods usually rely on the interval image shooting technique, which lacks continuity and might miss some critical moments during cracking process. This study proposes a deep learning-based video instance segmentation algorithm to locate, identify and segment soil cracks in real-time video stream. The training loss was found less than 0.1. The algorithm could record the cracks’ locations and numbers simultaneously. Besides, the crack ratio of clay could be calculated by crack pixels divided by total clay pixels among the entire soil cracking process. The proposed video instance segmentation method monitored the soil crack reached a number of 25 and the crack ration was about 8.5% at the experiment end, which has demonstrated the potential application for crack monitoring of geotechnical infrastructures via surveillance cameras.
