Neural Representation for 3D Building Reconstruction from Point Clouds

The continuous rise in urban growth has underscored the importance of airborne LiDAR point clouds (APCs) for efficient/cost effective urban planning, management, and development. However, the delineation and modeling of 3D objects -- specifically buildings -- from APCs pose significant challenges due to issues such as façade/roof occlusions, point density variations, sparsity, and noise. This thesis aims to address these challenges and provide neurally-driven solutions for 3D digital twinning of buildings from APCs. To delineate an urban scene into object categories, PReFormer, a memory-efficient point transformer capable of achieving competitive segmentation results with fewer model parameters and less memory is proposed. The PReFormer comprises of an optimized point embedding module, linearized multi-head self-attention layers, and reversible functions, all designed to reduce computation time and space complexities. Additionally, the architectural design of the PReFormer follows a ∇-shape, which improves (object) size-invariant feature extraction and segmentation accuracy. To generate high fidelity 3D building models from delineated APC building instances, APC2Mesh, a framework which integrates 3D building point completion and reconstruction processes, is proposed. The developed point completion network uses dynamic edge convolution and self-attention mechanism operations to extracts both local and global building shape information for complete building point set (BPS) reconstruction. This completion process mitigates the sparsity, occlusion, and point density variability issues usually associated airborne LiDAR BPSs. Leveraging the completed BPSs, a linearized skip-attention-based deformation network capable of handling several building styles and/complexities is presented to generate high fidelity 3D building mesh models. An observation of the mesh models from APC2Mesh shows that mesh models have relatively high disk storage, and are difficult to manipulate given their numerous triangular mesh faces. Thus, Points2Model, a neural-guided method that reconstructs building wireframes from APCs is further proposed. It uses neural implicit learning to up-sample completed BPSs, and a simple yet robust corner-focused hypothesis and selection strategy to detect building corners and their corresponding edge connectivity. Overall, this thesis presents innovative solutions for overcoming the inherent challenges of APCs in 3D building reconstruction, thus contributing significantly to the field of digital twinning of urban buildings.

Name thesis:degree_name

Doctor of Philosophy (PhD)

Discipline thesis:degree_discipline

Engineering – Geomatics

Grantor dc:publisher.institution

Graduate Studies

Year dc:date.issued

2024

Author dc:creator

• Akwensi, Perpetual Hope

Advisors dc:contributor.advisor

• Wang, Ruisheng
• Hassan, Quazi Khalid

Committee members dc:contributor.committeemember

• Ioannou, Yani Andrew
• Yang, Hongzhou

• LiDAR
• 3D Point Clouds
• Semantic Segmentation
• 3D Mesh Reconstruction
• 3D Wireframe Reconstruction
• 3D Point Completion
• Deep learning

Statement dc:rights

• University of Calgary graduate students retain copyright ownership and moral rights for their thesis. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission.

Language dc:language.iso

en

DOI

https://dx.doi.org/10.11575/PRISM/47452

OAI identifier oai:identifier

oai:ucalgary.scholaris.ca:1880/119841