A Survey on Point Cloud-based 3D Mesh Reconstruction for Large Scale Facilities

Digital twins replicate physical objects in virtual environments, enabling real-time monitoring and simulation. Their importance in industrial applications continues to grow with advances in AI technologies. A key requirement for digital twin construction is converting point clouds acquired from sensors such as LiDAR and depth cameras into reliable mesh models. Although preprocessing, registration, segmentation, and mesh reconstruction have been widely studied, most existing methods focus on small-scale point clouds. Processing large-scale point clouds containing hundreds of millions of points remains challenging because of computational and memory limitations. This paper reviews existing studies within a four-stage pipeline: preprocessing, registration, segmentation, and mesh reconstruction. Traditional and deep learning-based approaches are compared in terms of their strengths and limitations. Traditional methods are computationally efficient but often struggle with complex geometries because of local computation constraints. Deep learning methods capture global features and improve shape restoration, but their computational cost limits application to large-scale industrial datasets. Based on this analysis, we propose three future directions: (1) a hierarchical pipeline combining random sampling and 2D projection-based deep learning for large-scale point cloud processing, (2) an intelligent restoration strategy using component libraries and topological information to reconstruct missing regions, and (3) the construction of an industry-specific dataset and a standard-based verification framework for reliable engineering analysis.