摘要
本文针对盾构隧道三维激光点云数据,利用PointNet点云识别算法,构建隧道要素分割的深度神经网络模型。基于各要素样本数据与隧道壁背景数据存在类别不平衡情况,首先通过InstanceHardnessThreshold算法对隧道壁数据进行采样,平衡类别比例,同时借助focal loss函数对PointNet网络进行优化,使其更加着重于样本数目较少要素类别的学习,以提升分割效果。结果表明,本文研究方法能有效分割出横向施工缝、环向施工缝、螺栓孔等盾构隧道典型要素,可为隧道可视化和模型分析提供重要支撑。
A deep neural network model for tunnel element segmentation has been constructed in this paper with the 3 D laser point cloud data of shield tunnels and the PointNet point cloud recognition algorithm. Based on the classification imbalance between the sample data of each element and the background data of the tunnel wall, this paper has sampled the tunnel wall data through the InstanceHardnessThreshold algorithm to balance the classification ratio, and optimized the PointNet network with the focal loss function to make it more focused on learning on samples of small number to improve the segmentation effect. The results show that the research method in this paper can effectively segment the typical elements of shield tunnels such as transverse construction joints, circumferential construction joints, and bolt holes, and provide important support for tunnel visualization and model analysis.
作者
王兵海
WANG Binghai(China Railway Fifth Survey and Design Institute Group Co.Ltd.,Beijing 102600,China)
出处
《铁道建筑技术》
2022年第12期159-163,共5页
Railway Construction Technology
基金
中国铁建股份有限公司科技重大专项(2019-A02)。