高级检索

基于深度学习的大规模室外场景Scan2BIM研究综述

Deep Learning-based Scan2BIM for Large-Scale Scenes: a Comprehensive Review

  • 摘要: 当前, 大规模室外基础设施的数字化需求持续扩大, 基于深度学习的自动扫描到建筑信息模型(scanning to building information modeling, Scan2BIM)通过卓越的特征学习能力和自动化流程显著提升了建模精度和构建速度, 在结构复杂的室外场景重建中发挥了关键作用. 文中介绍了Scan2BIM的4大核心模块及其相关研究进展. 其中, 针对3D点云获取模块, 从采集设备与采集来源2个维度概括了3D点云数据采集的技术发展, 并着重梳理了代表性3D点云数据集; 根据学习方式的不同, 将大规模点云对齐算法划分为基于优化和深度学习2大类, 并从精准度、计算效率、鲁棒性等多维度对比分析了相关工作; 在点云分割模块中, 分别对点云全景分割和点云实例分割算法通过统一的评估指标进行了整理归纳; 对于BIM自动化建模, 简述了BIM核心互操作标准体系, 并分类总结了多种几何实体建模与关系建模算法. 最后, 通过深入分析和前瞻性探讨, 指出了现阶段大规模室外场景建模的高效性、精准性、泛化性与统一性的无法有效结合的问题; 未来将重点围绕多源数据融合建模、精度与鲁棒性协同优化、端到端Scan2BIM通用框架构建以及大模型应用与探索等方向展开.

     

    Abstract: With the increasing demand for digitizing large-scale outdoor infrastructure, the method of deep learning-based Scan2BIM (Scanning to Building Information Modeling) substantially improves modeling precision and construction speed by leveraging its powerful feature learning capabilities and automated workflows, which play a crucial role in the 3D reconstruction of structurally complex outdoor scenes. This paper introduces the four main modules of Scan2BIM and their corresponding research advancements. Firstly, the development of 3D point cloud data acquisition is summarized from the perspectives of both acquisition devices and sources, with dedicated emphasis on the systematic organization and analytical evaluation of representative 3D point cloud datasets. Secondly, large-scale point cloud registration algorithms are systematically categorized into two principal classes, according to learning approaches, which are optimization-driven and deep learning-based, followed by the multidimensional comparative analysis of existing methodologies through critical performance metrics including alignment precision, computational efficiency, and operational robustness. Thirdly, the algorithms of point cloud segmentation for both point cloud panoramic segmentation and point cloud instance segmentation are analyzed based on standardized evaluation metrics. Fourthly, the core BIM interoperability standard framework is briefly outlined, and various geometric entity and relational modeling algorithms are categorized in the automated BIM modeling module. Finally, through in-depth analysis and prospective discussion, this paper pointed out the challenges in achieving a balanced integration of efficiency, accuracy, generalization, and uniformity for large-scale outdoor scene modeling. Future work will primarily focus on multi-source data fusion modeling, the synergistic optimization of accuracy and robustness, the construction of end-to-end general Scan2BIM frameworks, and the application and exploration of large models.

     

/

返回文章
返回