Author Login Editor-in-Chief Peer Review Editor Work Office Work

Computer Engineering ›› 2020, Vol. 46 ›› Issue (12): 80-87. doi: 10.19678/j.issn.1000-3428.0056542

• Artificial Intelligence and Pattern Recognition • Previous Articles     Next Articles

A Robust Monocular Visual Odometry Algorithm

PANG Zhihua, QI Chenkun   

  1. School of Mechanical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
  • Received:2019-11-08 Revised:2019-12-18 Published:2020-01-03

一种鲁棒的单目视觉里程计算法

庞智华, 齐臣坤   

  1. 上海交通大学 机械与动力工程学院, 上海 200240
  • 作者简介:庞智华(1993-),男,硕士研究生,主研方向为视觉SLAM、多传感器融合;齐臣坤,副教授、博士生导师。
  • 基金资助:
    国家重点研发计划"政府间国际科技创新合作"重点专项(2017YFE0112200)。

Abstract: To address the low accuracy and robussness of the monocular visual odometry,this paper proposes a monocular visual odometry algorithm that combines the tree-based feature matching algorithm and the real-time map update strategy.The tree-based feature matching algorithm does not have motion assumptions,and can quickly and reliably establish feature matching relationships for various motions,ensuring the real-time performance and robustness of the algorithm.The real-time map update framework separates map expansion from optimization,which ensures the real-time performance of the algorithm and enables the current frame to track enough 3D map points for fast or violent motions,which greatly improves the robustness of the algorithm.In addition,in order to reduce the impact of real-time map updates on the accuracy of map points,a parallax-based weight matrix estimation method is proposed to make high-precision map points dominate the optimization functions to ensure the accuracy of the algorithm.The experimental results on two public datasets show that compared with the ORB-SLAM algorithm,the proposed algorithm has higher location accuracy,and has better robustness for fast or violent motions of the camera.

Key words: visual odometry, monocular vision, feature matching, real-time map update, pose optimization, robustness

摘要: 针对单目视觉里程计在相机快速或剧烈运动时精度差和鲁棒性低的问题,提出一种基于树结构特征匹配与实时地图更新策略的单目视觉里程计算法。基于树结构的特征匹配算法无运动假设,在各种运动下均可快速且可靠地建立特征匹配关系,从而保证算法的实时性和鲁棒性。实时地图扩展更新框架将地图扩展与优化分离,在保证实时性的同时又能使得当前帧在快速或剧烈运动下能跟踪到足够多的3D地图点,以提高算法的鲁棒性。为降低实时地图更新对地图点精度的影响,提出一种基于视差的权重矩阵估计方法,使高精度地图点在优化函数中占主导地位,以此保证算法的精度。在2个公开数据集上的实验结果表明,与ORB-SLAM算法相比,该算法定位精度更高,在相机快速或剧烈运动下具有更好的鲁棒性。

关键词: 视觉里程计, 单目视觉, 特征匹配, 实时地图更新, 位姿优化, 鲁棒性

CLC Number: