Semantically-rich as-built 3D modeling of the built environment from point cloud data

Modeling of the as-built environment is used by the Architecture, Engineering, Construction, and Facilities Management (AEC/FM) industry in a variety of engineering analysis scenarios. Significant applications include progress monitoring of construction sites, quality control of fabrication and on-site assembly, energy performance assessment, and structural integrity evaluation. The modeling process mainly consists of three sequential steps: data collection, modeling, and analysis. In current practice, these steps are performed manually by surveyors, designers, and engineers. Such manual tasks can be time-consuming, prohibitively expensive, and are prone to errors. While the analysis stage is fairly quick, taking several hours to complete, data collection and modeling can be the bottlenecks of the process: the first can spread over a few days, and the latter can span over multiple weeks or even months. Hence, the applicability of as-built modeling has been traditionally restricted to high latency analysis, where the model need not be updated frequently. In fast changing environments such as construction sites or tenant improvement projects, due to the difficulty in rapidly updating 3D models, model-based assessment methods for purposes such as progress or quality monitoring or creating the basis for design have had very limited applications. There is a need for a low-cost, reliable, and automated method for as-built modeling. This method should quickly generate and update accurate and complete semantically-rich models in a master format that is translatable to any engineering scenario and can be widely applied across all construction projects. This dissertation address these limitations by presenting an end-to-end procedure based on machine learning algorithms for generating semantically-rich 3D models from point cloud data. It specifically offers three new steps: (1) a new method that segments point clouds into semantically and geometrically meaningful subsets; this method outperforms the state-of-the-art solutions by addressing the problems of over-segmentation and semantic labeling beyond planar surfaces; (2) a new method that fits non-uniform b-spline curves and surfaces (NURBS) into the segmented point cloud subsets. This method also addresses several key manual steps within the state-of-the-art surface and NURBS fitting methods; and (3) a new method that reasons about the interconnectivity between NURBS surfaces based on formalized knowledge of construction sequencing and in particular geometrical relationship to connect the surfaces and finally represent semantically-rich 3D CAD models. My work also introduces a new dataset of point clouds from real-world commercial and industrial buildings. These point clouds were carefully modeled with cross-validation. A comprehensive set of machine learning validation techniques are used to present experimental results. The results show the end-to-end procedure can significantly low time and cost during a typical Scan to BIM process. The results also show that each individual method in the presented pipeline outperforms the most relevant state-of-the-art techniques. The benefits and limitations of the practice of Scan to BIM are discussed in detail.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Civil Engineering

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2020

Author dc:creator

• Perez, Yeritza

Contributors dc:contributor

• Golparvar-Fard, Mani
• El-Rayes, Khaled
• Liu, Liang
• El-Gohary, Nora
• Hoiem, Derek

• ScanToBIM
• Semantic Segmentation
• Geometric Segmentation
• Point Cloud
• Deep Learning
• Machine Learning
• Segmentation
• NURBS

Statement dc:rights

• Copyright 2020 Yeritza Perez

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/108716

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/108716