Automated Geometric Digital Twin Construction for Existing Buildings from Point Cloud Datasets

Abstract

The Architecture, Engineering, Construction, and Operation (AECO) industry significantly impacts the economy and environmental sustainability. However, it faces ongoing challenges in productivity and sustainability caused by outdated and inefficient information management methods. Digital Twins (DTs) offer a promising solution by enhancing decision-making processes throughout the building lifecycle. They have proven their value in all stages of the building lifecycle, including the operation and maintenance stages. However, most existing buildings lack DTs because they were constructed before the widespread adoption of digital technologies, while their design models are unreliable in representing as-is conditions of assets. Creating DTs requires capturing the as-is geometry of buildings (geometric DTs or gDTs). This process involves capturing Point Cloud Datasets (PCDs) and modelling these datasets to represent current building geometry accurately. However, this process requires extensive manual labour and remains a barrier to the broader adoption of DTs for the operation and maintenance of existing buildings. This thesis proposes a novel framework to automate the digitisation of building geometry from PCDs, comprising three main components: space detection, relation detection, and object detection. The space detection method uses empty blob detection and expansion to model volumetric spaces directly from large-scale PCDs, identifying walls, doorways, and windows. The relation detection method identifies primitive surfaces and establishes topological relations between them using line-casting and distance approaches followed by data-driven context-aware classification. This approach constructs a graph that interconnects building elements into a coherent structure. The object detection method combines data-driven and model-driven approaches and outcomes of the previous methods to detect and generate volumetric models for the most common building object types. Lastly, the proposed framework stitches the outcomes of individual steps together, also improving the consistency of the framework's outcome. Key findings demonstrate that the proposed method effectively detects and models spaces of arbitrary shapes from PCDs, including non-convex and complex ceiling structures. It outperforms existing methods for detecting Manhattan spaces and accurately identifies the locations of walls, doorways, and windows. The proposed method outperforms existing methods and achieves over 90 per cent of Intersection over Union for spaces from large-scale and occluded point clouds. The relation detection method successfully establishes geometric relations between objects and spaces, which improves object detection. By integrating distance-based and visibility-based methods, the framework effectively eliminates negative pairs of surfaces. We show that the proposed method archives 94.5% to 97.4% precision in detecting topological relations in indoor environments. Lastly, the object detection step generates volumetric objects and associated point clusters to represent the as-is geometry of individual objects. It leverages both data-driven and model-driven approaches to model volumetric objects in large-scale and occluded PCDs. The experiments show that the method archives over 90 per cent average precision in detecting most object types, which exceeds the performance of existing methods. The stitched outcomes of the detection of spaces, objects, and their relations allow for mutual adjustments, improving overall accuracy, such as in doorway geometry detection. The framework significantly reduces manual labour, saving approximately 90 per cent of the time required for manual modelling. In conclusion, this novel framework for automating geometry digitisation from PCDs addresses significant challenges in the AECO industry by enhancing productivity and accuracy in building information management. It improves the efficiency of constructing gDTs for existing buildings and ensures comprehensive and interconnected digital models. The implementation of this framework can lead to substantial time and cost savings, promoting broader adoption of DTs in the industry and contributing to more sustainable and well-managed building operations.