计算机工程 ›› 2019, Vol. 45 ›› Issue (12): 222-231,236.doi: 10.19678/j.issn.1000-3428.0052746

• 图形图像处理 • 上一篇    下一篇

基于概率运动统计特征匹配的单目视觉SLAM

曾维林, 刘桂华, 陈豪   

  1. 西南科技大学 信息工程学院, 四川 绵阳 621010
  • 收稿日期:2018-09-25 修回日期:2019-01-01 发布日期:2019-01-15
  • 作者简介:曾维林(1993-)男,硕士研究生,主研方向为图像处理、三维重建、SLAM算法;刘桂华,教授、博士;陈豪,本科生。
  • 基金项目:
    国防科工局核能开发科研项目"核应急处置机器人关键技术研究"([2016]1295);四川省科技厅重点研发项目"灵巧型激光自动导引运输装置(AGV)研制及其产业化"(17ZDYF0126)。

Monocular Vision SLAM Based on Probabilistic Motion Statistics Feature Matching

ZENG Weilin, LIU Guihua, CHEN Hao   

  1. School of Information Engineering, Southwest University of Science and Technology, Mianyang, Sichuan 621010, China
  • Received:2018-09-25 Revised:2019-01-01 Published:2019-01-15

摘要: 在单目视觉同步定位与建图(SLAM)过程中,由于特征匹配阶段存在误匹配且耗时长,使得机器人初始化速度慢、定位精度不高。针对此问题,基于概率运动统计特征匹配,提出一种单目视觉SLAM算法。通过设置自适应的阈值提取ORB特征点并使用四叉树进行保存,根据运动的平滑性与特征匹配的一致性估计特定区域内特征匹配的概率模型,得到正确的特征匹配点进行匹配,完成系统自动初始化与机器人位姿跟踪。在TUM数据集上的实验结果表明,该算法在特征匹配阶段耗时仅为1.4 ms,机器人初始化时间和定位精度分别为1.7 s和0.54 cm,且具有良好的实时性。

关键词: 单目视觉同步定位与建图, 概率运动统计特征匹配, ORB特征, 初始化, 定位精度

Abstract: In the current monocular vision Simultaneous Localization and Mapping(SLAM) process,the feature matching phase is time-consuming,and mismatches tend to occur,so the robot often has a slow initialization speed and low positioning accuracy.To address these problems,we propose a monocular vision SLAM algorithm based on probabilistic motion statistical feature matching.Firstly,we set the adaptive threshold to extract the ORB feature points,which are simultaneously saved by the quadtree.Then,according to the smoothness of motion and the consistency of feature matching,we estimate the probability model of feature matching in a specific region.Thereby,the correct feature matching points are obtained and the system automatically initialization and robot pose tracking are achieved subsequently.The experimental results on TUM dataset show that the algorithm only takes 1.4 ms in the feature matching phase.Besides,the robot initialization time and positioning accuracy are 1.7 s and 0.54 cm respectively,which represents a good real-time performance.

Key words: monocular vision Simultaneous Localization and Mapping(SLAM), probabilistic motion statistics feature matching, ORB features, initialization, positioning accuracy

中图分类号: