测绘学报 ›› 2018, Vol. 47 ›› Issue (6): 825-832.doi: 10.11947/j.AGCS.2018.20170619

• 计算机视觉与三维重建 • 上一篇    下一篇

三维点云数据实时管理的Hash map方法

郑顺义1, 何源1, 徐刚2, 王辰1, 朱锋博1   

  1. 1. 武汉大学遥感信息工程学院, 湖北 武汉 430079;
    2. 立命馆大学理工学部情报学科, 日本 滋贺 520072
  • 收稿日期:2017-11-06 修回日期:2018-04-08 出版日期:2018-06-20 发布日期:2018-06-21
  • 通讯作者: 何源 E-mail:yuanhe@whu.edu.cn
  • 作者简介:郑顺义(1973-),男,博士,教授,博士生导师,主要从事数字摄影测量与计算机视觉研究。E-mail:syzheng@whu.edu.cn
  • 基金资助:
    国家自然科学基金(41671452;41701532);中央高校基本科研业务费专项资金(2042016kf0012);中国博士后科学基金(2017M612510)

Hash Map Method of 3D Point Cloud Data for Real-time Organizing

ZHENG Shunyi1, HE Yuan1, XU Gang2, WANG Chen1, ZHU Fengbo1   

  1. 1. School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China;
    2. Department of Computer Science & Media Technology, Ritsumeikan University, Shiga 520072, Japan
  • Received:2017-11-06 Revised:2018-04-08 Online:2018-06-20 Published:2018-06-21
  • Supported by:
    The National Natural Science Foundation of China (Nos.41671452;41701532);The Fundamental Research Funds for the Central Universities (No.2042016kf0012);China Postdoctoral Science Foundation (No.2017M612510)

摘要: 本文基于机器视觉探讨数字摄影测量三维构像下的智能数据处理要素之一:海量点云高效管理技术,提出了一种基于GPU的hash map三维点云数据组织的改进算法,算法可以高效地完成数据的动态插入、更新和索引,而不受数据规模限制。同时,通过传感器位置姿态估计当前活动范围,进行主机与GPU的数据交换,保证了GPU的低内存占用率。在搭载不同等级显卡(GTX960、GTX1050、GTX1060)的计算机设备上试验,本文算法均可以达到60 fps以上的帧率(单帧处理点云数:2.11×105),证明算法满足了三维构像中三维点云数据高效管理的要求。

关键词: 三维点云, 数据组织, 哈希表, 图形处理器, 并行计算

Abstract: In this paper,one of the key elements of intelligent data processing in digital photogrammetry is discussed based on machine vision:efficient management technology of massive point cloud,an improved algorithm for hash mapping 3D data on GPU is proposed.The algorithm can efficiently perform dynamic insertion,update and indexing of data without the limitation of data scale.We applied the algorithm to TSDF (truncated signed distance field) algorithm for point cloud fusion,which can reduce the noise of single frame and the data registration error between different frames in a very efficient way.Currently,most of the data structures of point cloud fusion algorithms are regular or hierarchical grid data structures,so the object bounding boxes should be specified in advance.Moreover,the hierarchical data structure is complex and hard to parallelization.Therefore,the requirements of dynamic scalability,updating and indexing data of real-time 3D reconstruction cannot be satisfied.This paper exploited hash map to manage 3D data.At the same time,the current active region could be estimated with sensor motion to exchange data between host and GPU,keeping low memory utilization of GPU.In the experiments on different levels of graphics cards(GTX960,GTX1050,GTX1060),the algorithm can satisfy the real-time frame rate requirements(60 fps,2.11×105 points per frame),so it meets the requirement of efficient management of 3D point cloud data in three-dimensional imaging.

Key words: 3D point cloud, data organization, hash table, graphics processing unit (GPU), parallel computing

中图分类号: