Research on 3D Reconstruction Methods for Incomplete Building Point Clouds Using Deep Learning and Geometric Primitives

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Point cloud data, known for their accuracy and ease of acquisition, are commonly used for reconstructing level of detail 2 (LoD-2) building models. However, factors like object occlusion can cause incompleteness, negatively impacting the reconstruction process. To address this challenge, this paper proposes a method for reconstructing LoD-2 building models from incomplete point clouds. We design a generative adversarial network model that incorporates geometric constraints. The generator utilizes a multilayer perceptron with a curvature attention mechanism to extract multi-resolution features from the input data and then generates the missing portions of the point cloud through fully connected layers. The discriminator iteratively refines the generator’s predictions using a loss function that is combined with plane-aware Chamfer distance. For model reconstruction, the proposed method extracts a set of candidate polygons from the point cloud and computes weights for each candidate polygon based on a weighted energy term tailored to building characteristics. The most suitable planes are retained to construct the LoD-2 building model. The performance of this method is validated through extensive comparisons with existing state-of-the-art methods, showing a 10.9% reduction in the fitting error of the reconstructed models, and real-world data are tested to evaluate the effectiveness of the method.
Funding Information
  • National Natural Science Foundation of China (42071419, GFZX0404130304, 2023ZBSK041, 23WL(Y)53)
  • High-resolution Earth Observation System of China (42071419, GFZX0404130304, 2023ZBSK041, 23WL(Y)53)
  • Zibo City Social Science Planning Research Project of China (42071419, GFZX0404130304, 2023ZBSK041, 23WL(Y)53)
  • Shandong Province Culture and Tourism Research Project of China (42071419, GFZX0404130304, 2023ZBSK041, 23WL(Y)53)

This publication has 42 references indexed in Scilit: