运动曲率自适应V-SLAM算法研究  

Motion curvature adaptive V-SLAM algorithm research

作　　者：赵荣亮 王红旗[1] 刘群坡[1] 张岩 ZHAO Rongliang;WANG Hongqi;LIU Qunpo;ZHANG Yan(School of Electrical Engineering and Automation,Henan Polytechnic University,Jiaozuo,Henan 454000,China)

机构地区：[1]河南理工大学电气工程与自动化学院,河南焦作454000

出　　处：《导航定位学报》2022年第6期53-58,共6页Journal of Navigation and Positioning

摘　　要：传统室内视觉同时定位与建图(V-SLAM)存在特征点误匹配剔除耗时长、较大转弯时定位精度低等问题。为了提升匹配速度,提出一种基于特征点的余弦相似度直方图剔除误匹配方法。为了避免大幅度转弯时关键帧丢失,提出一种基于运动曲率的关键帧自适应提取V-SLAM算法(MSAV-SLAM)。该算法通过高效多点透视(EPnP)计算帧间位姿,以融合了运动曲率、帧间位移和旋转矩阵的复合运动量为参考变量,自适应选择关键帧。室内数据集实验表明,所提出的算法与基于固定时间、空间和原ORB-SLAM2算法相比,平均每帧追踪时间减少为原算法的57.62%,且绝对轨迹误差为改进前的81.45%。Traditional indoor vision simultaneous localization and mapping(V-SLAM) has many problems, such as timeconsuming feature point mismatch elimination and low localization accuracy in large turns. In order to improve the matching speed, a method of eliminating false matching based on cosine similarity histogram of feature points is proposed. In order to avoid the loss of key frames in large turns, a motion state adaptive key frame extraction V-SALM(MSAV-SLAM) algorithm was proposed. This algorithm calculates the interframe pose by efficient perspective-n-point(EPnP), and adaptively selects key frames by taking the combined motion curvature, interframe displacement and the compound motion of rotation matrix as reference variables. Laboratory data set experiments show that compared with the original ORB-SLAM2 algorithm based on fixed time and space, the average tracking time per frame of the proposed algorithm is reduced to 57.62%, and the absolute trajectory error is 81.45%.

关 键 词：视觉同步定位与地图创建 关键帧 特征点匹配 运动曲率 室内定位 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]