文献类型：期刊文献

中文题名：基于最短路径的大视角变换下SLAM回环检测

英文题名：SLAM Loop Closure Detection Based on the Shortest Path Under Perspective Distortions

作者：应桃[1];严怀成[1];吴滢跃[2];王孟[1];李郅辰[1]

机构：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237;[2]中国科学院上海技术物理研究所,上海200083

年份：2021

卷号：28

期号：11

起止页码：2158

中文期刊名：控制工程

外文期刊名：Control Engineering of China

收录：CSTPCD;;北大核心:【北大核心2020】；CSCD:【CSCD_E2021_2022】；

基金：国家自然科学基金资助项目(62073143,61922063);上海学术带头人项目(19XD1421000);上海港澳台国际合作项目(19510760200);上海教委科技创新项目(2021-01-07-00-02-E00107)。

语种：中文

中文关键词：视觉SLAM;回环检测;大视角变换;最短路径算法;图像协方差矩阵

外文关键词：Visual simultaneous localization and mapping(SLAM);loop closure detection;perspective distortion;the shortest path algorithm;image covariance matrix

摘要：回环检测是视觉同步定位与地图构建中必不可少的一部分,回环的正确检测能够避免机器人的累计位姿漂移问题。但是移动机器人在定位过程中,相机视角发生大幅变化会使得相邻帧间图像的特征点匹配数目骤减,匹配精度降低,系统性能下降甚至匹配失效。基于这些关键帧进行回环检测时会因为视角的变换使得原本的相同位置误认为不同位置,减少了正确回环的判断结果,降低了系统的召回率,影响准确的地图构建。因此,提出基于图像协方差矩阵匹配的最短路径算法,以关键帧为节点,帧间协方差矩阵距离为权重,构建有向图,通过最短路径算法得到关键帧间的位置关系,再结合距离阈值确定相似位置,完成回环检测。最后,在数据集和实际场景中验证了这一算法。
Loop closure detection is an essential part of visual simultaneous location and mapping. Correctly detecting loop closures can help mobile robots reduce the accumulated pose drift. However, during the location process of the mobile robot, when the camera’s angle of view changes significantly, the number of feature points matching between adjacent frames will be sharply reduced, the matching accuracy will be reduced, and the system performance will decline or even fail. When loop closure detection is performed based on these key frames, the same location will be mistaken due to the change of perspective distortions, which reduces the accuracy of the loop closure. It reduces the recall rate of the system and affect the accurate map construction.Therefore, the shortest path algorithm based on image covariance matrix matching is proposed, which uses key frames as nodes and inter-frame covariance matrix distance as weights to construct a directed map. The location relationship between the key frames is obtained, and the similar location is determined by the distance threshold. Finally, the algorithm is verified with data set and actual scenes.