• CN:11-2187/TH
  • ISSN:0577-6686

机械工程学报 ›› 2021, Vol. 57 ›› Issue (2): 219-229.doi: 10.3901/JME.2021.02.219

• 交叉与前沿 • 上一篇    下一篇

扫码分享

二进制点云局部特征描述子研究

唐敏杰, 赵欢, 丁汉   

  1. 华中科技大学数字制造装备与技术国家重点实验室 武汉 430074
  • 收稿日期:2020-04-22 修回日期:2020-10-05 出版日期:2021-01-20 发布日期:2021-03-15
  • 通讯作者: 赵欢(通信作者),男,1983年出生,教授。主要研究方向为机器人智能化加工装备与技术。E-mail:huanzhao@hust.edu.cn
  • 作者简介:唐敏杰,男,1995年出生。主要研究方向为智能传感技术与机器人视觉。E-mail:tamgminjie@hust.edu.cn;丁汉,男,1963年出生,教授,博士研究生导师。主要研究方向为数字化制造与机器人技术。E-mail:dinghan@hust.edu.cn
  • 基金资助:
    国家自然科学基金(52090054)和湖北省自然科学基金杰青(2020CFA077)资助项目。

Research on Binarized Local Feature Descriptors of Point Clouds

TANG Minjie, ZHAO Huan, DING Han   

  1. State Key Laboratory of Digital Manufacturing Equipment and Technology, Huazhong University of Science and Technology, Wuhan 430074
  • Received:2020-04-22 Revised:2020-10-05 Online:2021-01-20 Published:2021-03-15

摘要: 基于点云局部特征描述的三维目标识别是机器人视觉领域一个具有重要研究价值且富有挑战性的研究方向。尽管目前已有大量三维特征描述子的相关研究工作,但它们大多数采用浮点数,对计算和存储的开销很大,并且鉴别力较弱,鲁棒性不强。鉴于此,从点对特征出发,提出一种鉴别力高,鲁棒性强,结构紧凑,计算迅速的高性能点云局部描述算法—二进制点对特征直方图(Binarized histogram of point pair features,B-HPPF)。对模型进行降采样,根据点位置与点法线信息,计算局部邻域中点对的七个特征;利用其将局部点对集划分为若干区域,并对每一区域进行信息提取;通过轮换比较各信息量的大小将特征进行二进制编码;将每一区域的二进制子特征串联组合生成最终的二进制描述子B-HPPF。所提出的B-HPPF描述子在多个公开数据集上进行测试,并与经典的描述算法进行对比,结果表明,所提出的方法在鉴别力、鲁棒性、紧凑性和计算效率等方面获得了优越的综合性能。此外,B-HPPF的实用性也在目标识别数据集上得以进一步验证。

关键词: 点云, 局部特征, 特征提取, 二值化, 目标识别

Abstract: 3D object recognition based on local feature description of point cloud is a challenging problem in the robotic. Although a large number of three-dimensional feature descriptors have been proposed, most of them use floating-point numbers, which are expensive for calculation and storage. In view of this, a high-performance three-dimensional local descriptor is proposed called binarized histogram of point-pair feature(B-HPPF). After downsampling the model, based on the point position and point normal information, extract seven features of each point pair in the local neighborhood; use these features to divide the local point pair set into several regions and information is ex-tracted. The features are binary-coded by comparing the size of the corresponding information. The binary sub-features of each region are combined in series to generate a final binary B-HPPF description. The B-HPPF descriptors are tested on a number of public datasets and compared with classical description algorithms. The results show that it is optimally balanced in terms of discriminative power, robustness, compactness and computational effi-ciency. Moerover, B-HPPF is applied to the target identification data set to verify its practicability.

Key words: point cloud, local feature, feature extrction, binarized, object recogntion

中图分类号: