尺寸质量检测是成品房屋在交付前的必要步骤，但传统人工检测方法耗时费力。随着利用陆地激光扫描仪的自动化尺寸质量检测得到更多关注，室内多视角点云自动化配准变得更加重要。由于在室内布置标靶的效率偏低，且成品房屋室内充斥着大量重复的结构，不适合仅依赖自然几何基元或俯视图进行无标靶配准。本文提出了一种基于深度学习的室内多视角点云自动化配准方法。首先利用PointAF神经网络对扫描点云数据进行语义分割，再进行实例分割得到不同结构的点云实例。随后，利用门实例进行两两配准计算变换参数，通过基于重叠置信度和冲突约束的评价函数去除错误匹配，并使用基于生成树的顺序配准方法完成多视角配准。在验证和对比实验中，利用2套成品房屋共21站扫描点云数据证明了所提出方法的有效性和精度。

Dimensional quality inspection is a necessary step for finished residences before delivery, but traditional manual inspection methods are time-consuming and labor-intensive. As automated dimensional quality inspection utilizing terrestrial laser scanners receives more attention, automated multi-view point cloud registration of indoor scenes becomes more important. Due to the fact that posting targets indoors is inefficient and a large number of repetitive structures fill the indoor scenes of finished residence, it is not suitable to rely solely on natural geometric primitives or top views for target-less registration. In this paper, a deep learning-based automated multi-view point cloud registration method for indoor scenes is proposed. Firstly, the PointAF neural network is utilized to semantically segment the scanned point cloud data, and then instance segmentation is performed to obtain point cloud instances with different structures. Subsequently, pairwise registration is performed to compute the transformation parameters using the door instances. False matches are removed by an evaluation function based on overlapping confidence and conflict constraints. Finally, multi-view registration is achieved using a spanning tree based sequential registration method. In the validation and comparison experiments, a total of 21 stations of scanned point cloud data from 2 sets of finished residences are utilized to demonstrate the effectiveness and accuracy of the proposed method.

国家自然科学基金(52130801、52108283)