地球信息科学学报 ›› 2019, Vol. 21 ›› Issue (4): 615-622.doi: 10.12082/dqxxkx.2018.0452

• 论文 • 上一篇    下一篇

基于区域特征分割的密集匹配点云渐进形态学滤波

张刚1,2(), 刘文彬2, 张男2   

  1. 1. 中国测绘科学研究院,北京 100830
    2. 北京四维远见信息技术有限公司,北京 100039
  • 收稿日期:2018-09-10 修回日期:2019-03-25 出版日期:2019-04-24 发布日期:2019-04-24
  • 作者简介:

    作者简介:张 刚(1974-),男,吉林长春人,高级工程师,主要从事摄影测量及应用研究。E-mail: 43131904@qq.com

  • 基金资助:
    国家重点研发计划项目(2017YFB0503004);高分辨率对地观测系统重大专项(42-Y2-0A14-9001-17/18)

Progressive Morphological Filtering Method of Dense Matching Point Cloud based on Region Feature Segmentation

Gang ZHANG1,2,*(), Wenbin LIU2, Nan ZHANG2   

  1. 1. Chinese Academy of Surveying and Mapping, Beijing 100830, China
    2. Beijing Geo-Vision Technology Company Limited, Beijing 100039, China
  • Received:2018-09-10 Revised:2019-03-25 Online:2019-04-24 Published:2019-04-24
  • Contact: Gang ZHANG E-mail:43131904@qq.com
  • Supported by:
    National Key Research and Development Program of China, No.2017YFB0503004;National S&T Major Project for High-resolution Earth Observation System, No.42-Y2-0A14-9001-17/18

摘要:

随着计算机视觉和遥感技术的进步,基于遥感影像的密集匹配也成为目前获取高精度点云的重要手段之一。与LiDAR点云类似,点云数据处理的基础步骤就是点云滤波。在数据特征上,密集匹配生成的点云与LiDAR获取的点云既类似但又有区别。本文在渐进形态学滤波算法上添加了特征条件,将点云和图像结合成深度图像,并对深度图像按典型地物类型进行语义分割,从而对与图像平面坐标一致的点云进行标记和首次滤波;然后按几何特征将场景简单分类,按分类结果对应的参数滤波构建地面点三角网;最后综合初滤波结果和语义分割类型标记对特征相似的区域进行优化确认,得到最终的滤波结果,并与布料模拟滤波(CSF)算法进行了对比验证实验。结果表明,基于特征的渐进形态学滤波其I类误差在1.98%以内,Ⅱ类误差在2.33%以内,较适宜对精度要求较高的应用,尤其是混合地形的滤波。

关键词: 密集匹配, 点云滤波, 布料滤波, 深度学习, 区域特征分割, 渐近形态学滤波, 无人机

Abstract:

With the progress of computer vision and RS, dense matching based on remote sensing images has also become one of the important means to obtain high-precision point clouds. Like point clouds of LiDAR, filtering is the fundamental step. Dense matching point cloud is similar with LiDAR point cloud, but have different feature. In this paper, the feature condition is added to the progressive morphological filtering algorithm, point clouds and images are combined into RGB-Depth images, and depth images are semantically segmented according to typical object types, so that point clouds which coordinate correspond with image coordinate are marked and filtered for the first time. Then divide point clouds by grid, then do simply classified according to geometric features, and the improved irregular triangular network of ground points is constructed by filter parameters corresponding to the classification results. Finally, use and intergraded the pre-filtering results and the semantic segmentation results, the regions with similar features are optimized and confirmed by predefined parameter, and the final filtering results are obtained. The results are compared with results of the Cloth Simulated Filtering algorithm. The test result was show that type I error less than 1.98%, type II error less than 2.33% of the progressive morphological filtering algorithm, that algorithm is suitable for higher precision application, especially mixed terrain points cloud filtering.

Key words: dense matching, point cloud filtering, cloth simulation, deep learning, region feature segmentation, morphological filtering, UAV