PLVO: Plane-line-based RGB-D visual odometry

PLVO: Plane-line-based RGB-D visual odometry

Abstract

A plane-line-based RGB-D visual odometry (PLVO) is proposed to solve the degenerate problem in the pose estimation of an RGB-D camera using plane features. First, the plane-line hybrid association graph (PLHAG) is proposed to associate two types of geometric features. Planes and lines are associated in an integrated framework, which fully exploits the geometric relationships between two planes and between a plane and a line, respectively. Then, the pose of an RGB-D camera is estimated based on the adaptive fusion of planes and lines. Generally speaking, the plane features are more accurately and stably extracted than the line features. As a result, in our method, the planes dominate the calculation of the camera pose through an adaptive weighting algorithm. As for the degrees of freedom (DoFs) of the pose that cannot be constrained by planes, the line features are supplementarily used to obtain the full 6 DoF pose estimation of the camera. Thus, the fusion of two types of features is achieved and the degenerate problem using only plane features is solved. Various experiments on public benchmarks as well as in real-world environments demonstrate the efficiency of the proposed method.

Key words: RGB-D visual odometry / plane and line fusion / robot localization / adaptive fusion / multi-feature joint association

摘要

针对利用平面特征计算RGB-D相机位姿时的求解退化问题, 提出平面和直线融合的RGB-D视觉里程计(Plane-line-based RGB-D visual odometry, PLVO). 首先, 提出基于平面−直线混合关联图(Plane-line hybrid association graph, PLHAG)的多特征关联方法, 充分考虑平面和平面、平面和直线之间的几何关系, 对平面和直线两类几何特征进行一体化关联. 然后, 提出基于平面和直线主辅相济、自适应融合的RGB-D相机位姿估计方法. 具体来说, 鉴于平面特征通常比直线特征具有更好的准确性和稳定性, 通过自适应加权的方法, 确保平面特征在位姿计算中的主导作用, 而对平面特征无法约束的位姿自由度(Degree of freedom, DoF), 使用直线特征进行补充, 得到相机的6自由度位姿估计结果, 从而实现两类特征的融合, 解决了单纯使用平面特征求解位姿时的退化问题. 最后, 通过公开数据集上的定量实验以及真实室内环境下的机器人实验, 验证了所提出方法的有效性.

关键词: RGB-D视觉里程计 / 平面−直线融合 / 机器人定位 / 自适应融合 / 多特征联合关联

Recommended citation: 孙沁璇, 苑晶, 张雪波, 高远兮. PLVO: 基于平面和直线融合的RGB-D视觉里程计. 自动化学报, 2023, 49(10): 2060−2072 doi: 10.16383/j.aas.c200878 http://sunqinxuan.github.io/files/publications-2023-10-24-AAS.pdf