计算机工程

所属专题: 智能机器人专题

• 智能机器人专题 • 上一篇    下一篇

基于局部与全局优化的双目视觉里程计算法

杨冬冬1,2,张晓林1,李嘉茂1   

  1. (1.中国科学院上海微系统与信息技术研究所,上海 200050; 2.中国科学院大学,北京 100049)
  • 收稿日期:2017-01-13 出版日期:2018-01-15 发布日期:2017-01-15
  • 作者简介:杨冬冬(1991—),男,硕士研究生,主研方向为计算机视觉、机器人导航;张晓林,研究员;李嘉茂,副研究员。
  • 基金项目:

    国家自然科学基金青年项目(61601448)。

Binocular Visual Odometry Algorithm Based on Local and Global Optimization

YANG Dongdong 1,2,ZHANG Xiaolin 1,LI Jiamao 1   

  1. (1.Shanghai Institute of Microsystem and Information Technology,Chinese Academy of Sciences,Shanghai 200050,China;2.University of Chinese Academy of Sciences,Beijing 100049,China)
  • Received:2017-01-13 Online:2018-01-15 Published:2017-01-15

摘要:

为实现移动机器人的实时精确定位,提出一种新的双目视觉里程计算法。利用加速尺度不变特征变换(SIFT)算子提取左右图像特征并做稀疏立体匹配,对前后帧图像进行帧间SIFT特征追踪,在RANSAC策略下通过运动估计获得初始位姿(旋转和平移矩阵)。在此基础上,将图像序列分为关键帧和非关键帧,采用可变滑动窗口对相邻关键帧的位姿局部非线性优化以减小帧间运动估计误差,同时通过词袋模型进行回环检测,对环内所有关键帧的位姿全局优化,避免位姿误差的累积和轨迹漂移。实验结果表明,该算法满足实时性要求,并且能够减小位姿误差,提高定位精度。

关键词: 双目视觉里程计, 运动估计, 局部优化, 全局优化, 特征匹配

Abstract:

A novel binocular Visual Odometry(VO) algorithm is proposed for real-time precise localization of mobile robots.Firstly,it uses accelerated Scale Invariant Feature Transform(SIFT) operator to extract the image features on the left and right image.The sparse stereo matching is carried out after the extracting.In addition,the method of features tracking is applied between the previous and current image.Thus,the initial pose including rotation and translation matrix can be obtained with the motion estimation method based on the RANSAC strategy.Secondly,the image sequence is divided into key frames and non-key frames.In order to decrease the error of the inter-frame motion estimation,a variable sliding window is applied to optimizing the pose of adjacent key frames locally and nonlinearly.Finally,the closed-loop detection is applied by the method of bag of words.Furthermore,all the poses of key frames in the closed-loop are optimized globally to avoid the error accumulation and the drift of the trajectory.Experimental results show that the proposed algorithm has good real-time performance,while reducing the position pose error and improving the positioning accuracy.

Key words: binocular Visual Odometry(VO), motion estimation, local optimization, global optimization, feature matching

中图分类号: