PRIMITIVE-BASED BUILDING MODEL RECONSTRUCTION FROM POINT CLOUDS

Modern remote sensing technologies, such as airborne photogrammetry and lidar, provide extensive point cloud data for reconstructing 3D building models in large urban areas. While existing model-driven and data-driven building reconstruction approaches offer effective solutions, they each have inherent limitations. Data-driven methods are sensitive to the source data and often struggle with generalization to avoid geometric inconsistencies, whereas model-driven methods require well-designed roof primitives and massive labeled data for training. To address these challenges, this thesis introduces a parametric roof primitive library and proposes three innovative reconstruction approaches: single primitive reconstruction, multi-primitive reconstruction, and hybrid reconstruction, each tailored for different building structures and complexities.

The single primitive reconstruction method focuses on accurately identifying and reconstructing both the primary structures and superstructures of buildings, ensuring a detailed and comprehensive representation. The multi-primitive reconstruction approach extends this capability to more complex buildings by decomposing them into multiple roof primitives and employing an iterative RANSAC-based fitting process to determine their parameters. While these methods achieve precise segmentation through a novel holistic fitting technique, they encounter challenges in handling highly irregular structures and large-scale datasets due to computational constraints and the limited availability of predefined roof primitives. To address these issues, the hybrid reconstruction approach integrates a learning-based method trained on a label-free synthetic dataset for roof primitive identification and parameter determination. For building components that cannot be represented using predefined primitives, Regularized Explicit Polyhedral (REP) models are applied, guided by a holistic polygonal regularization technique, enhancing the adaptability and scalability of the reconstruction process.

Extensive evaluations on airborne and spaceborne photogrammetric and lidar datasets validate the effectiveness of these approaches. The proposed methods achieve high horizontal and vertical reconstruction accuracy, generating semantically and geometrically coherent building models. Their ability to process multi-source point clouds underscores their versatility in urban modeling applications. Beyond theoretical advancements, these approaches have practical implications in urban planning, infrastructure monitoring, and virtual city modeling, offering valuable tools for smart city development, disaster management, and augmented reality simulations. By introducing scalable, robust, and adaptive methodologies, this research advances the field of 3D building reconstruction, paving the way for more automated and efficient urban modeling solutions.