首页 >  , Vol. , Issue () : -

摘要

全文摘要次数: 170 全文下载次数: 243
引用本文:

DOI:

10.11834/jrs.20211009

收稿日期:

2021-01-07

修改日期:

2021-06-25

PDF Free   EndNote   BibTeX
3D点云震害建筑物深度学习样本增强方法
摘要:

摘要:针对震后复杂场景LiDAR点云建筑物破坏类型的自动识别问题,为满足应急救援时效性、准确性需求,告别传统人工震害特征提取,充分挖掘点云数据中灾区建筑物震害信息,进一步实现建筑物自动化智能化识别。本文将3D点云深度学习方法应用于建筑物震害识别,构建了包含倒塌、局部倒塌、未倒塌三种建筑物破坏类型的点云数据集。基于PointNet++网络探究了各类别样本量及其均衡性对识别精度的影响,并提出破坏建筑物样本增强方法,丰富了各类别样本点云形态。利用2010年海地7.0级地震后机载LiDAR数据,在PointNet++网络中进行了样本增强前后分类精度比较、样本量以及均衡性分析实验,样本增强后倒塌和局部倒塌的分类精度分别提高近27%和17%,模型整体平均分类精度、Kappa系数均有近15%的提升。实验结果表明三维建筑物震害深度学习模型在各类别样本量足够且均衡时,才能取得较好的分类识别效果。

Research on Enhancement Method of Point Cloud Seismic Damaged Building Samples Based on PointNet++
Abstract:

Objective Aiming at the problem of automatic identification of building damage types in post-earthquake complex lidar point cloud scenes, and in order to meet the requirements of timeliness and accuracy of emergency rescue, bid farewell to traditional artificial seismic damage feature extraction, fully excavate the seismic damage information of buildings in the disaster area in the point cloud data, and further realize the automatic and intelligent recognition of buildings.A building seismic damage recognition model was established based on the PointNet++ network, and three point cloud training datasets including collapsed, partially collapsed, and uncollapsed were constructed, which is a scientific earthquake Emergency rescue and disaster assessment provide an important basis. Method This paper attempts to apply the 3D point cloud deep learning method to the identification of seismic damaged buildings. And due to the uneven sample size, based on the characteristics of the PointNet++ network and the original point cloud sample shape , we propose a sample enhancement method including inverse distance interpolation, symmetry and top projection to increase the amount of collapsed and partial collapsed samples . Result After sample enhancement, it not only increases the number of collapsed buildings and partial collapsed buildings, making the samples more comprehensive and diverse, but also solves the problem of uneven samples。Therefore, the classification accuracy of collapse and partial collapse is improved by about 30% and 20%, and the overall average classification accuracy and kappa coefficient of the model are improved by more than 10%. The difference in classification accuracy between collapsed and uncollapsed, partial collapsed and uncollapsed are reduced from 40% and 30% to about 15%. Conclusion (1) The characteristics of all kinds of point cloud samples and network model should be fully considered when building seismic damage training data set. In this paper, we consider that PointNet + + has the same learning characteristics for the same geometric shape in different scales and spatial rotation changes, and design a sample enhancement method for the collapse and partial collapse category, which not only increases the number of samples, but also enriches the damage form of samples, and effectively improves the classification accuracy of local collapse. (2) The sample size and sample size balance have a greater impact on the recognition effect of the earthquake damage recognition model established by the PointNet++ network. Only when the sample size is sufficient and the number of samples in each category is relatively uniform, can a better classification and recognition effect be achieved. However, the sample size is not the decisive factor for improving the classification effect. When the sample size is uniform, the uncollapsed accuracy is still higher than the other two categories. This is also related to factors such as sample selection, network design, and internal feature learning methods, which can be further explored. Key words: classification and recognition; PointNet++; sample enhancement; seismic damage buildings; LiDAR point cloud

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群 分享按钮