In recent years, monitoring the health condition of existing bridges has become a common requirement. By providing an information management system, Bridge Information Model (BrIM) can highly improve the efficiency of health inspection and the reliability of condition evaluation. However, the current modeling processes still largely rely on manual work, where the cost outweighs the benefits. The main barrier lies in the challenging step of semantic segmentation of point clouds. Efforts have been made to identify and segment the structural components of bridges in existing research. But these methods are either dependent on manual data preprocessing or need big training dataset, which, however, has rendered them unpractical in real-world applications. This paper presents a combined local descriptor and machine learning based method to automatically detect structural components of bridges from point clouds. Based on the geometrical features of bridges, we design a multi-scale local descriptor, which is then used to train a deep classification neural network. In the end, a result refinement algorithm is adopted to optimize the segmentation results. Experiments on real-world reinforced concrete (RC) slab and beam-slab bridges show an average precision of 97.26%, recall of 98.00%, and intersection over union (IoU) of 95.38%, which significantly outperforms PointNet. This method has provided a potential solution to semantic segmentation of infrastructures by small sample learning and will contribute to the fulfillment of the automatic BrIM generation of typical highway bridges from the point cloud in the future.

近年来，监测既有桥梁的健康状况已成为普遍要求。通过提供信息管理系统，桥梁信息模型（BrIM）可以极大地提高健康检查的效率和状态评估的可靠性。然而，目前的建模过程仍然很大程度上依赖于手工工作，成本大于收益。主要障碍在于点云语义分割的挑战性步骤。在现有的研究中，已经努力识别和分割桥梁的结构部件。但是这些方法要么依赖于手动数据预处理，要么需要大的训练数据集，然而，这使得它们在实际应用中不切实际。本文提出了一种结合局部描述符和基于机器学习的方法，以从点云中自动检测桥梁的结构组件。基于桥梁的几何特征，我们设计了一个多尺度局部描述符，然后用于训练深度分类神经网络。最后采用结果细化算法对分割结果进行优化。在真实世界的钢筋混凝土 (RC) 板和梁板桥上的实验表明，平均精度为 97.26%，召回率为 98.00%，联合交集 (IoU) 为 95.38%，明显优于 PointNet。