[1]曹一波,赵鹏飞,朱海文,等.结构化环境下基于点线面特征融合的 SLAM 算法[J].计算机技术与发展,2023,33(07):85-90.[doi:10. 3969 / j. issn. 1673-629X. 2023. 07. 013]
　CAO Yi-bo,ZHAO Peng-fei,ZHU Hai-wen,et al.SLAM Algorithm Based on Point-line-plane Feature Fusion in Structured Environment[J].,2023,33(07):85-90.[doi:10. 3969 / j. issn. 1673-629X. 2023. 07. 013]

点击复制

结构化环境下基于点线面特征融合的 SLAM 算法()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

33

期数:

2023年07期

页码:

85-90

栏目:

移动与物联网络

出版日期:

2023-07-10

文章信息/Info

Title:

SLAM Algorithm Based on Point-line-plane Feature Fusion in Structured Environment

文章编号:

1673-629X(2023)07-0085-06

作者:

曹一波; 赵鹏飞; 朱海文; 刘 顺; 张智辉

华南师范大学 软件学院,广东 佛山 528200

Author(s):

CAO Yi-bo; ZHAO Peng-fei; ZHU Hai-wen; LIU Shun; ZHANG Zhi-hui

School of Software,South China Normal University,Foshan 528200,China

关键词:

点线面特征; SuperPoint; 同时定位与地图构建; 结构化环境; 重投影误差

Keywords:

point-line-plane feature; SuperPoint; SLAM; structured environment; reprojection error

分类号:

TP242. 6+2

DOI:

10. 3969 / j. issn. 1673-629X. 2023. 07. 013

摘要:

结构化场景中,存在着低纹理表面为特征的人造环境,基于点特征的 SLAM ( Simultaneous Localization and Mapping,同时

定位与地图构建) 算法难以得到足够的匹配点对,从而导致相机估计运动失败。 除了点之外,结构化环境提供了大量的几何

特征,例如线和平面。 因此,提出一种基于点线面特征融合的 SLAM 算法。 算法将基于深度学习的 SuperPoint 点特征与传统线面特征相结合,利用结构化场景的特性,将位姿解耦细化。 首先,使用线面特征构建 MW( Manhattan World,曼哈顿世界) 坐标系,利用每一时刻相机与 MW 坐标系的相对旋转得到相机之间的旋转矩阵;然后,构建点线面特征的重投影误差函数,通过最小化联合误差函数得到平移矩阵;最后,根据结构化环境下平面间相互垂直和平行的特性添加约束函数,同时为弥补环境中出现不严格遵守 MW 假设的情况,使用关键帧构建的局部地图投影到当前帧进一步优化位姿。 在TUM 公开数据集上与主流方法对比表明,该算法有效提升了结构化低纹理环境下的 SLAM 定位精度。

Abstract:

In structured scenes, there are artificial environments characterized by low textured surfaces, so it is difficult for the SLAM( Simultaneous Localization and Mapping) algorithm based on point feature to get enough matching point pairs,which leads to the failureof camera motion estimation. In addition to points,structured environments provide a number of geometric features,such as lines andplanes. Therefore,a SLAM algorithm based on point,line and surface feature fusion is proposed. The algorithm combines the SuperPointfeatures based on deep learning with the traditional line-plane features,and utilizes the features of structured scenes to refine the positionand pose decoupling. First, the?

MW ( Manhattan World) coordinate system is constructed by using line and plane features, and therotation matrix between the cameras is obtained by using the relative rotation between the cameras and the MW coordinate system?

at eachmoment. Then,the reprojection error function of point,line and plane features is constructed,and the translation matrix is obtained byminimizing the joint error function. Finally,constraint functions are added according to the vertical and parallel characteristics betweenplanes in the structured environment. In order to compensate for the non-strict compliance with the MW hypothesis in the environment,the local map constructed by key frames is projected to the current frame to further optimize the pose. On TUM open dataset, theproposed algorithm can effectively improve the localization accuracy of SLAM in structured low-texture environment compared with themainstream methods.

常用功能