Automated Pipe Spool Recognition in Cluttered Point Clouds
MetadataShow full item record
Construction management is inextricably linked to the awareness and control of 3D geometry. Progress tracking, quality assurance/quality control, and the location, movement, and assembly of materials are all critical processes that rely on the ability to monitor 3D geometry. Therefore, advanced capabilities in site metrology and computer vision will be the foundation for the next generation of assessment tools that empower project leaders, planners, and workers. 3D imaging devices enable the capture of the existing geometric conditions of a construction site or a fabricated mechanical or structural assembly objectively, accurately, quickly, and with greater detail and continuity than any manual measurement methods. Within the construction literature, these devices have been applied in systems that compare as-built scans to 3D CAD design files in order to inspect the geometrical compliance of a fabricated assembly to contractually stipulated dtolerances. However, before comparisons of this type can be made, the particular object of interest needs to be isolated from background objects and clutter captured by the indiscriminate 3D imaging device. Thus far, object of interest extraction from cluttered construction data has remained a manual process. This thesis explores the process of automated information extraction in order to improve the availability of information about 3D geometries on construction projects and improve the execution of component inspection, and progress tracking. Specifically, the scope of the research is limited to automatically recognizing and isolating pipe spools from their cluttered point cloud scans. Two approaches are developed and evaluated. The contributions of the work are as follows: (1) A number of challenges involved in applying RANdom SAmple Consensus (RANSAC) to pipe spool recognition are identified. (2) An effective spatial search and pipe spool extraction algorithm based on local data level curvature estimation, density-based clustering, and bag-of-features matching is presented. The algorithm is validated on two case studies and is shown to successfully extract pipe spools from cluttered point clouds and successfully differentiate between the specific pipe spool of interest and other similar pipe spools in the same search space. Finally, (3) the accuracy of curvature estimation using data collected by low-cost range-cameras is tested and the viability of use of low-cost range-cameras for object search, localization, and extraction is critically assessed.