• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 图形与图像 • 上一篇    下一篇

基于深度学习特征点法的单目视觉里程计

熊炜1,2,金靖熠1,王娟1,刘敏1,曾春艳1   

  1. (1.湖北工业大学电气与电子工程学院,湖北 武汉 430068;
    2.南卡罗来纳大学计算机科学与工程系,哥伦比亚 29201)
     
  • 收稿日期:2019-07-13 修回日期:2019-08-30 出版日期:2020-01-25 发布日期:2020-01-25
  • 基金资助:

    国家自然科学基金(61571182,61601177);湖北省自然科学基金(2019CFB530);国家留学基金(201808420418)

Monocular visual odometry based on
deep learning feature point method

XIONG Wei1,2,JIN Jing-yi1,WANG Juan1,LIU Min1,ZENG Chun-yan1   

  1. (1.School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China;
    2.Department of Computer Science and Engineering,University of South Carolina,Columbia SC 29201,USA)
     
  • Received:2019-07-13 Revised:2019-08-30 Online:2020-01-25 Published:2020-01-25

摘要:

针对特征点法的视觉里程计VO中光度、视点变化对特征点提取稳定性降低的不利影响,提出一种基于深度学习特征点法的单目VO方法。采用自监督深度学习网络训练得到DSP特征点检测器。首先使用亮度非线性逐点调整方法对训练图像进行光度调整;然后使用非极大值抑制方法剔除冗余DSP特征点,改进最邻近方法得到双向最邻近方法,解决特征点匹配问题;最后建立最小化重投影误差方程求解优化位姿及空间点参数。在Hpatches、Visual Odometry数据集上进行验证,实验结果表明:DSP特征点检测器增强了特征匹配对光度、视点变化的鲁棒性;无后端优化的条件下,本方法定位均方根误差比ORB方法明显降低,且保证了系统实时性,为特征点法的VO提供新的解决思路。

关键词: 视觉里程计, 深度学习, 亮度非线性逐点调整, 特征匹配, 重投影误差

Abstract:

Aiming at the adverse effect of luminosity and viewpoint change on feature point extraction stability in Visual Odometry (VO) of feature point method, a monocular VO method based on deep learning feature point method is proposed. The deep learning SuperPoint (DSP) feature point detector is obtained by self-supervised deep learning network training. Firstly, the brightness of the training image is adjusted by the brightness nonlinear point-by-point adjustment method. Secondly, the redundant DSP feature points are eliminated by using the non-maximum value suppression method. The two-way nearest neighbor algorithm is improved based on the nearest neighbor algorithm to solve the feature point matching problem. Finally, the equation for minimizing the reprojection error is established to solve the optimal pose and spatial point parameters. The experimental results on Hpatches and Visual Odometry datasets show that the DSP feature point detector enhances the robustness of feature matching to luminosity and viewpoint changes. Without the backend optimization, this method reduces the root mean square error obviously in comparison to ORB method. The real-time performance of the system is guaranteed, which provides a new solution for the VO of feature-based method.
 

Key words: Visual Odometry (VO), deep learning, brightness nonlinear point-by-point adjustment, feature matching, reprojection error