Towards automated processing of 3D point clouds for the generation of geometric-semantic models of existing buildings

The fields of Architecture, Engineering, Construction and Operation (AECO) have and continue to experience a fundamental paradigm shift towards digitalization. Not only do digital models of buildings and civil structures provide methods for describing their geometric shape, they also offer additional semantic information about individual components, thus giving them an edge over paper-based plans and documentation. In this context, Building Information Modelling (BIM) encompasses concepts for digital models which transparently represent buildings throughout their entire life and methods for creating these models either during planning or by deriving them from existing documentation or on-site survey data during the operation phase. BIM consequently draws its advantages from digitalisation paradigms like the simplified data exchange between stakeholders to improve communication and software revolving around automated analysis, with asset maintenance and management built on top of it. Specifically the oftentimes manual creation of BIM models representing an existing asset's as-is state poses a major hurdle in the widespread adoption of BIM and is even more daunting when considering the large number of objects constructed long before the introduction of BIM. However, recreating this as-is state either requires paper plans which are usually not available or severely outdated, therefore only leaving accurate on-site measurements for the creation of up-to-date models as an option. The advent of 3D reality capturing techniques such as laser scanning and Dense Image Matching introduced the necessary means for capturing of such real-world 3D geometries as point clouds. Recent innovations have given rise to various new techniques such as mobile laser scanning which benefit from increased flexibility and faster capturing and low-cost sensors with similar properties. These trends have fueled the creation of Scan-to-BIM workflows which aim to partially or completely automate the creation of geometric-semantic models. While cutting down on processing time and costly human intervention is the main motivator behind this development, the degree of automation and the degree of user intervention require careful balancing and thus complicate this process. This is especially apparent when considering that different stages are involved in Scan-to-BIM, which are built on top of one another. Following the concept of Scan-to-BIM, this thesis deals with the automated creation of BIM models from point cloud data and therefore explores all stages of the related workflow. These are in sequential order: capturing and preprocessing of captured point clouds, segmentation, classification of segments, and reconstruction of the final BIM as-is model. Due to all steps representing a set of closely interwoven problems, methods combining them are not uncommon and will be presented in the articles contained in this cumulative thesis. For the preprocessing stage two works are presented, the first of which being meant for processing point clouds right after capturing. Said method reorders the point cloud and then decomposes it into wavelet frequencies to generate a Level of Detail representation capable of continuous interpolation which can aid in inspection and downsampling. The wavelet representation additionally lends itself to both, lossless and lossy compression and can thus reduce the on-disk footprint of point cloud data. The second presented method directly impacts subsequent stages of the Scan-to-BIM pipeline, as it aims to align point clouds alongside a global axis. This transformation notably benefits voxel-based methods further down the processing pipeline by preventing aliasing artefacts from manifesting and therefore improves their robustness. A collection voxel-based methods covering all remaining stages of Scan-to-BIM for indoor scenarios are shown afterwards, where all of them are combined into a workflow allowing for fast spatial segmentation of individual storeys and rooms, with a final wall reconstruction step for BIM model fitting. To overcome the limited voxel resolution, all stages of the workflow refine their results by using the underlying points to provide both, high precision and processing speed. Aside from earlier methods which are based purely on classical point cloud processing algorithms, machine learning has earned its place in segmentation and classification workflows as well. A workflow using deep learning-based point cloud classification is presented in the final publication. As a way of circumventing the limited amount of point cloud training data, the approach uses a convolutional neural network for semantic image segmentation and projects the resulting segmentation masks into the point cloud. Obtained labels are then postprocessed using classical methods such as majority-voting, point cloud region growing and feature-based filtering to refine results. The published works and their multi-disciplinary nature show that point cloud processing as part of Scan-to-BIM is a complex problem which can be tackled from various perspectives. As proven by the presented implementations, the modular nature of such workflows naturally lends itself to being supplemented with new techniques which promise high geometrical accuracy and additional semantic information and thus points towards future research.