Towards Reverse Engineering of Industrial Site Plants

Abstract

CAD models of industrial sites are extremely important, as they provide documentation and simplify inspection, planning, modification, as well as a variety of physical and logistics simulations of the corresponding installations. Despite these clear advantages, many industrial sites do not have CAD models, or have trouble keeping them up-to-date. This is often due to the amount of effort required to create and maintain CAD models updated. Hopefully, the recent popularization of 3D scanning devices is promoting the development of reverse engineering, allowing the creation of 3D representations of real environments from point clouds. Nevertheless, point clouds extracted from industrial sites are extremely complex due to occlusions, noise, non-uniform sampling, size of the dataset, lack of sample organization, among other factors. Thus, a successful reverse engineering solution should have several desirable properties, including speed, robustness to noise, accuracy, and be able to handle point clouds in general without requiring one to fine tune their parameters to each dataset in order to work well on it. This thesis presents some initial efforts towards obtaining a robust framework for reverse engineering of industrial sites. It introduces two fast and robust algorithms for detecting, respectively, planes and cylinders in noisy unorganized point clouds. Planes and cylinders are typically the most common and largest structures found in those environments, representing walls, floors, ceilings, pipes, and ducts. We demonstrate the effectiveness of the proposed approaches by comparing their performances against the state-of-the-art solutions for plane and cylinder detection in unorganized point clouds. In these experiments, our solutions achieved the best overall accuracy using the same set of (default) parameter values for all evaluated datasets. This is in contrast to the competing techniques, for which their parameter values were individually adjusted for each combination of technique and dataset to achieve their best results in each case, demonstrating the robustness of our algorithms, which do not require fine-tuning to perform well on arbitrary point clouds. Moreover, our technique also displayed competitive speed to other state-of-art techniques, being suitable for handling large-scale point clouds. The thesis also presents a graphical user interface which allows further refinement of the detected structures, providing the user the ability to remove, merge, and semi-automatically detect planes and cylinders in point clouds.