PointCrack3D: Crack Detection in Unstructured Environments using a\n 3D-Point-Cloud-Based Deep Neural Network

Abstract

Surface cracks on buildings, natural walls and underground mine tunnels can\nindicate serious structural integrity issues that threaten the safety of the\nstructure and people in the environment. Timely detection and monitoring of\ncracks are crucial to managing these risks, especially if the systems can be\nmade highly automated through robots. Vision-based crack detection algorithms\nusing deep neural networks have exhibited promise for structured surfaces such\nas walls or civil engineering tunnels, but little work has addressed highly\nunstructured environments such as rock cliffs and bare mining tunnels. To\naddress this challenge, this paper presents PointCrack3D, a new\n3D-point-cloud-based crack detection algorithm for unstructured surfaces. The\nmethod comprises three key components: an adaptive down-sampling method that\nmaintains sufficient crack point density, a DNN that classifies each point as\ncrack or non-crack, and a post-processing clustering method that groups crack\npoints into crack instances. The method was validated experimentally on a new\nlarge natural rock dataset, comprising coloured LIDAR point clouds spanning\nmore than 900 m^2 and 412 individual cracks. Results demonstrate a crack\ndetection rate of 97% overall and 100% for cracks with a maximum width of more\nthan 3 cm, significantly outperforming the state of the art. Furthermore, for\ncross-validation, PointCrack3D was applied to an entirely new dataset acquired\nin different locations and not used at all in training and shown to detect 100%\nof its crack instances. We also characterise the relationship between detection\nperformance, crack width and number of points per crack, providing a foundation\nupon which to make decisions about both practical deployments and future\nresearch directions.\n