Research Institute for Smart Cities & Shenzhen Key Laboratory of Spatial Information Smart Sensing and Services, School of Architecture and Urban Planning, Shenzhen University, Shenzhen 518060, China.
State Key Laboratory of Information Engineering in Surveying Mapping and Remote Sensing, Wuhan University, Wuhan 430000, China.
Sensors (Basel). 2019 Jan 27;19(3):533. doi: 10.3390/s19030533.
Semantically rich indoor models are increasingly used throughout a facility's life cycle for different applications. With the decreasing price of 3D sensors, it is convenient to acquire point cloud data from consumer-level scanners. However, most existing methods in 3D indoor reconstruction from point clouds involve a tedious manual or interactive process due to line-of-sight occlusions and complex space structures. Using the multiple types of data obtained by RGB-D devices, this paper proposes a fast and automatic method for reconstructing semantically rich indoor 3D building models from low-quality RGB-D sequences. Our method is capable of identifying and modelling the main structural components of indoor environments such as space, wall, floor, ceilings, windows, and doors from the RGB-D datasets. The method includes space division and extraction, opening extraction, and global optimization. For space division and extraction, rather than distinguishing room spaces based on the detected wall planes, we interactively define the start-stop position for each functional space (e.g., room, corridor, kitchen) during scanning. Then, an interior elements filtering algorithm is proposed for wall component extraction and a boundary generation algorithm is used for space layout determination. For opening extraction, we propose a new noise robustness method based on the properties of convex hull, octrees structure, Euclidean clusters and the camera trajectory for opening generation, which is inapplicable to the data collected in the indoor environments due to inevitable occlusion. A global optimization approach for planes is designed to eliminate the inconsistency of planes sharing the same global plane, and maintain plausible connectivity between the walls and the relationships between the walls and openings. The final model is stored according to the CityGML3.0 standard. Our approach allows for the robust generation of semantically rich 3D indoor models and has strong applicability and reconstruction power for complex real-world datasets.
语义丰富的室内模型在整个设施的生命周期中越来越多地用于不同的应用。随着 3D 传感器价格的降低，从消费级扫描仪获取点云数据变得非常方便。然而，由于视线遮挡和复杂的空间结构，大多数现有点云 3D 室内重建方法都涉及繁琐的手动或交互式过程。本文利用 RGB-D 设备获取的多种类型的数据，提出了一种从低质量的 RGB-D 序列快速自动构建语义丰富的室内 3D 建筑模型的方法。我们的方法能够从 RGB-D 数据集中识别和构建室内环境的主要结构组件，例如空间、墙壁、地板、天花板、窗户和门。该方法包括空间划分和提取、开口提取和全局优化。对于空间划分和提取，我们不是根据检测到的墙壁平面来区分房间空间，而是在扫描过程中交互式定义每个功能空间（例如，房间、走廊、厨房）的起始和停止位置。然后，提出了一种用于墙壁组件提取的内部元素过滤算法和用于空间布局确定的边界生成算法。对于开口提取，我们提出了一种新的基于凸壳、八叉树结构、欧式聚类和相机轨迹特性的噪声鲁棒性方法，用于开口生成，由于室内环境中不可避免的遮挡，该方法不适用于采集的数据。设计了一种用于平面的全局优化方法，以消除具有相同全局平面的平面之间的不一致性，并保持墙壁和墙壁与开口之间的关系之间的合理连接。最终模型根据 CityGML3.0 标准存储。我们的方法允许生成语义丰富的 3D 室内模型，并且对复杂的真实世界数据集具有很强的适用性和重建能力。
