测绘学报 ›› 2023, Vol. 52 ›› Issue (2): 260-271.doi: 10.11947/j.AGCS.2023.20210332

• 摄影测量学与遥感 • 上一篇    下一篇

点云场景语义标注的排序批处理模式主动学习法

邹禄杰1,2,3, 花向红1,2, 赵不钒1,2, 陶武勇1,2,4, 李琪琪1,2,3   

  1. 1. 武汉大学测绘学院, 湖北 武汉 430079;
    2. 武汉大学灾害监测与防治研究中心, 湖北 武汉 430079;
    3. 广州市城市规划勘测设计研究院, 广东 广州 510060;
    4. 南昌大学信息工程学院, 江西 南昌 330031
  • 收稿日期:2021-06-28 修回日期:2022-05-18 发布日期:2023-03-07
  • 通讯作者: 花向红 E-mail:xhhua@sgg.whu.edu.cn
  • 作者简介:邹禄杰(1996-),男,硕士生,研究方向为点云数据处理。E-mail:2015301610030@whu.edu.cn
  • 基金资助:
    国家自然科学基金(41674005;41871373)

Ranked batch-mode active learning method for semantic annotation of point cloud scene

ZOU Lujie1,2,3, HUA Xianghong1,2, ZHAO Bufan1,2, TAO Wuyong1,2,4, LI Qiqi1,2,3   

  1. 1. School of Geodesy and Geomatics, Wuhan University, Wuhan 430079, China;
    2. Research Center for Hazard Monitoring and Prevention, Wuhan University, Wuhan 430079, China;
    3. Guangzhou Urban Planning & Design Survey Research Institute, Guangzhou 510060, China;
    4. School of Information Engineering, Nanchang University, Nanchang 330031, China
  • Received:2021-06-28 Revised:2022-05-18 Published:2023-03-07
  • Supported by:
    The National Natural Science Foundation of China (Nos. 41674005;41871373)

摘要: 针对点云场景语义标注存在着手工标注费时费力、算法耗时严重、标注精度不高和不适用于大规模场景点云处理等问题,本文提出了一种结合排序批处理模式的主动学习点云场景语义标注方法。该方法首先对原始点云进行下采样处理,然后利用改进的递归特征增加法从庞大的特征集中筛选出最优特征子集,采用排序批处理模式采样算法迭代选取并人工标注少数未标注点,通过创建最小人工标注训练集来完成下采样点云的语义标注工作,最后利用邻域等权标签传播算法完成原始点云数据的标注。对3个室外大场景点云分别进行的试验表明:本文方法只需人工标注7.50%、7.35%、5.83%的点云即可完成下采样点云的标注工作。此外,对比试验表明,本文方法在标注精度和减少人工成本方面优于其他方法,能为点云语义标注工作节省大量人工成本。

关键词: 点云场景, 排序批处理模式, 特征选择, 主动学习, 语义标注

Abstract: Due to the semantic annotation of point cloud scene by manual annotation is time-consuming and label cost process, and the annotation accuracy is not high, the point cloud processing is not suitable for large-scale scenes, this paper proposed an active learning annotation method of point cloud based on ranked batch-mode. This method firstly downsampling the original point cloud, then an improved recursive feature addition method is used to filter out the optimal feature subset from a huge feature set, and a ranked batch-mode sampling algorithm is adopted to iteratively select and manually label fraction of unlabeled points. The semantic annotation of the down-sampled point cloud is completed by creating a minimum manual annotation training set, and finally the original point cloud data is annotated using the neighborhood equal-weight label propagation algorithm. Experiments on three outdoor large scene point clouds show that the method in this paper only needs to manually label 7.50%, 7.35%, and 5.83% of the point clouds to complete the labeling of the down-sampled point clouds. In addition, comparative experiments show that this method is superior to other methods in labeling accuracy and reducing labor costs, and save a lot of labor costs for point cloud semantic annotation work.

Key words: point cloud scene, ranked batch-mode, feature selection, active learning, semantic annotation

中图分类号: