基于环形特征匹配的双目视觉里程计  

Stereo Visual Odometry Based on Ring Feature Matching

作　　者：黄平[1] 曹镇 王欢[1] Huang Ping;Cao Zhen;Wang Huan(College of Intelligent Systems Science and Engineering,Harbin Engineering University,Harbin,Heilongjiang 150001,China)

机构地区：[1]哈尔滨工程大学智能科学与工程学院,黑龙江哈尔滨150001

出　　处：《光学学报》2021年第15期158-168,共11页Acta Optica Sinica

基　　金：国家自然科学基金(61871143);黑龙江省自然科学基金(LH2019F006);哈尔滨市应用技术研究与开发项目(2017R-AQXJ095)。

摘　　要：针对视觉里程计中采用基本特征点匹配算法易出现误匹配现象,提出一种结合双向光流法的特征点环形匹配算法。该算法将双目图像与前后帧图像组成环形,前后帧图像采用双向金字塔光流法跟踪特征点,并剔除错误的特征点追踪结果。基本特征匹配算法通常采用快速最近邻匹配(FLANN)算法,但其结果包含很多误匹配的点对。本文提出的匹配算法不仅能剔除误匹配特征点,还能使特征点均匀分布于图像上。随后采用基于3D-2D点的P3P(Perspective-3-Point)算法结合RANSAC(Random Sample Consensus)得到初始位姿估计结果,再利用g2o(General Graph Optimization)库对位姿估计结果进一步优化。定位实验结果表明,结合双向光流法的特征点环形匹配算法具有更高的定位精度。In this paper, regarding the phenomenon that the basic feature point matching algorithm is prone to mismatch in visual odometry, we proposes a ring matching algorithm of feature points combined with the bidirectional optical flow method. This algorithm forms a ring structure between the stereo image and the images in the front and rear frames. For the images in the front and rear frames, the bidirectional pyramid optical flow method is used to track feature points and eliminate mismatched feature points. The basic feature point matching algorithm usually adopts fast library for approximate nearest neighbors(FLANN), but the result contains many mismatched point pairs. The proposed matching algorithm can not only eliminate the mismatched feature points but also make the feature points evenly distributed on the images. Subsequently, the perspective-3-point(P3 P) algorithm based on 3 D-2 D points is combined with random sample consensus(RANSAC) to obtain the initial pose estimation results. The general graph optimization(g2 o) library is employed to further optimize the pose estimation results. The positioning experiments verify that the ring matching algorithm of feature points combined with the bidirectional optical flow method has higher positioning accuracy.

关 键 词：机器视觉 视觉里程计 环形特征匹配 位姿估计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]