摘要
为了解决现有语义分割研究对机载激光点云几何信息利用不足的问题,文中在Point Net++的基础上提出了一种基于几何注意力机制的机载激光点云语义分割方法。几何注意力层将低维几何特征向量作为先验,通过自注意力的方式获得高维局部式样的表示,从而获得具有几何判别性的点云深度特征。在Vaihingen数据集上取得了83.4%总体精度指标和70.4%的平均F1分数。在平均F1分数上,相较于Point Net++,RandLA-Net分别高出4.8%和2.7%。在DALES数据集中取得了97.8%的总体精度和81.1%的平均F1分数。在平均F1分数上,相较于Point Net++、RandLA-Net分别提升了12.8%和20.5%。实验结果表明,几何注意力层能有效提高点云地物类别区分的能力。
In order to solve the problem of insufficient utilization of the geometric information of airborne laser point cloud in existing research on semantic segmentation,this paper,on the basis of Point Net++,proposes a method for the semantic segmentation of airborne laser point cloud based on geometric attention mechanism.With the low dimensional geometric feature vector as a prior,the geometry attention layer obtains the high dimensional local pattern representation through self attention,thereby obtaining geometrically discriminative point cloud depth features.It achieves 83.4%overall accuracy and 70.4%average F 1 score on the Vaihingen dataset.The average F 11 score is 4.8%and 2.7%,which are higher than those of Point Net++and RandLA Net.The overall accuracy and average F 1 score reach 97.8%and 81.1%on the DALES dataset.Compared with Point Net++and RandLA Net,the average F 1 score increases by 12.8%and 20.5%respectively.The experimental results show that the geometric attention module can effectively improve the ability of point cloud object category distinction.
作者
赵佳楠
周少辉
王振
ZHAO Jianan;ZHOU Shaohui;WANG Zhen(Branch Company in Guangzhou,China Railway First Bureau Group Co.,Ltd,Guangzhou 511492)
出处
《西安工业大学学报》
CAS
2023年第2期180-188,共9页
Journal of Xi’an Technological University
基金
深圳西改扩建项目(2022A-063)。
关键词
机载激光点云
语义分割
几何注意力机制
深度学习
ALS point clouds
semantic segmentation
geometric attention mechanism
deep learning
