27 JAN 2026 (TUE) 15:05 - 15:35

Three Dimensional Building Modeling with LiDAR Point Clouds using Deep Learning 

Miss SUI Jialu   

( Supervisor: Prof Hongsheng Zhang )

Abstract:

Urban-scale three-dimensional building modeling constitutes a critical technical underpinning for optimizing urban energy allocation, boosting socioeconomic benefits, refining urban planning and governance systems, and advancing smart city development. As a core pillar of smart city construction, Building Information Modeling (BIM) is an integrated paradigm encompassing building geometric data, attribute information, and lifecycle management. It plays a pivotal role in urban planning, construction management, infrastructure operation, and digital twin development, laying the foundation for data-driven urban governance. Urban building three-dimensional modeling serves as the fundamental premise of urban-scale Building Information Modeling implementation and underpins various Building Information Modeling-enabled downstream applications. However, transitioning from single-building to large-scale urban Building Information Modeling faces significant bottlenecks, demanding robust three-dimensional modeling methodologies aligned with Building Information Modeling’s multi-dimensional requirements. 

With advancing urbanization, demand for building-scale digital twins has surged, driving research on 3D reconstruction focusing on façades, roofs and full structures. LiDAR point clouds from Airborne LiDAR Scanning (ALS) and Unmanned Aerial Vehicle (UAV) remain the primary data source for outdoor modeling due to their high precision. Yet, large urban environments characterized by diverse building shapes and close proximity, along with unique BIM application requirements, make urban building 3D modeling highly challenging. 

Key research gaps in integrating urban 3D modeling with BIM include: 1) Difficulties in accurate building instance segmentation from LiDAR point clouds in large urban areas, which serves as a prerequisite for individual BIM models; 2) LiDAR point cloud issues such as sparsity, non-uniform density, occlusion and noise hinder precise geometry reconstruction that meets BIM quality standards; 3) A rigid, uniform modeling paradigm fails to balance geometric detail, computational efficiency and data storage optimization across diverse BIM application scenarios including urban planning, construction execution and facility operation; 4) The disconnect between modeling outputs that only contain geometric data and BIM’s attribute and lifecycle information limits the scalability of urban BIM. 

To address the aforementioned bottlenecks in large-scale urban BIM implementation, this study proposes a BIM-oriented urban building 3D modeling framework focusing on three core objectives. 1) Develop and optimize building-scene-specific instance segmentation algorithms to establish an efficient, accurate pipeline integrating preprocessing and segmentation for raw LiDAR data, enabling precise extraction of individual building point clouds. 2) Propose a novel shape completion model for partial LiDAR point clouds to generate high-precision, detail-preserving 3D data that captures complex urban buildings’ structural topology and fine features, supporting high-accuracy reconstruction. 3) Explore key 3D geometric and topological characteristics of segmented building point clouds that guide reconstruction strategies, establish a corresponding 3D reconstruction framework, and verify its accuracy and applicability in complex urban scenarios.