关键词: 视觉SLAM, 曼哈顿世界, 低纹理环境, ICL-NUIM数据集, TUM数据集, 点线面特征, 绝对轨迹误差, 相机运动解耦

Abstract:

To solve the challenges faced by visual SLAM(Simultaneous Localization And Mapping) systems in low-texture environments, this study proposes an RGB-D SLAM algorithm based on point, line and plane features to enhance localization accuracy. The proposed algorithm is built upon the ORB-SLAM2(Oriented FAST and Rotated BRIEF-Simultaneous Localization and Mapping 2) framework and introduces the Manhattan world assumption to decouple camera pose into rotation and translation matrices, effectively mitigating the issue of error accumulation. In the aspect of feature extraction, ORB feature points and LSD(Line Segment Detector) algorithm are used to extract line features, and hierarchical clustering algorithm is used to extract plane features, making full use of the geometric information of spatial structure. The experimental results show that compared with the ORB-SLAM2 algorithm, the proposed algorithm performs better in multiple low-texture scenes in TUM and ICL-NUIM datasets. By comparing the root-mean-square error of the absolute trajectory error, the proposed algorithm significantly improves the positioning accuracy in low-texture environments, and has significant advantages in scenes with fewer feature points.

Key words: visual SLAM, Manhattan world, low-texture environments, ICL-NUIM dataset, TUM dataset, point-line-plane features, absolute trajectory error, camera motion decoupling