With the gaining popularity and proliferation of building information modeling (BIM) techniques, a growing demand emerges for accurate, up-to-date and semantically-enriched digital representations of built environments. In this regard, current mobile indoor mapping systems like the Microsoft HoloLens or Matterport allow to efficiently acquire triangle meshes of indoor building environments. However, manually reconstructing digital models of building interiors on the basis of these triangle meshes is a cumbersome and time-consuming task. Consequently, in this work, we propose a fully automatic, voxel-based indoor reconstruction approach to derive semantically-enriched and geometrically completed indoor models in voxel representation from unstructured triangle meshes. The presented approach does not require room surfaces such as walls, ceilings or floors to be planar or aligned with the coordinate axes. Furthermore, it does not rely on a clear vertical subdivision in distinct floor levels and even allows for slanted floors such as ramps or stair flights. It thus can also be applied to challenging indoor environments featuring curved room surfaces and complex vertical room layouts. The proposed approach labels voxels as ‘Ceiling’, ‘Floor’, ‘Wall’, ‘Wall Opening’, ‘Interior Object’ and ‘Empty Interior’. Room surfaces are geometrically completed in case of holes in the input triangle meshes caused by occlusion or incomplete mapping. Furthermore, the derived interior space is partitioned into rooms and connecting transition spaces. To demonstrate the performance of our approach, we conduct a thorough quantitative evaluation on four labeled benchmark datasets. To this aim, we present a novel and adequate, automatic evaluation method. The four datasets have been acquired with the Microsoft HoloLens and are available along with the manually modeled ground truth. We also release the code of our implementation of the voxel-based indoor reconstruction approach presented in this paper as well as the code for the automated evaluation against the ground truth data at https://github.com/huepat/voxir.

随着建筑信息建模 (BIM) 技术的日益普及和扩散，对建筑环境的准确、最新和语义丰富的数字表示的需求不断增长。在这方面，当前的移动室内地图系统（如 Microsoft HoloLens 或 Matterport）允许有效获取室内建筑环境的三角形网格。然而，基于这些三角形网格手动重建建筑物内部的数字模型是一项繁琐且耗时的任务。因此，在这项工作中，我们提出了一种全自动的、基于体素的室内重建方法，以从非结构化三角形网格的体素表示中导出语义丰富和几何完整的室内模型。所提出的方法不需要房间表面，例如墙壁，天花板或地板是平面的或与坐标轴对齐。此外，它不依赖于不同楼层的清晰垂直细分，甚至允许倾斜地板，例如坡道或楼梯。因此，它还可以应用于具有弯曲房间表面和复杂垂直房间布局的具有挑战性的室内环境。提议的方法将体素标记为“天花板”、“地板”、“墙壁”、“墙壁开口”、“内部对象”和“空的内部”。如果由于遮挡或不完整映射导致输入三角形网格中出现孔洞，则房间表面在几何上完成。此外，衍生的室内空间被划分为房间和连接过渡空间。为了证明我们方法的性能，我们对四个标记的基准数据集进行了彻底的定量评估。为了这个目标，我们提出了一种新颖且充分的自动评估方法。这四个数据集是通过 Microsoft HoloLens 获得的，可与手动建模的地面实况一起使用。我们还在 https://github.com/huepat/voxir 上发布了我们实现本文中提出的基于体素的室内重建方法的代码以及针对地面实况数据的自动评估代码。