Robust RGB-D Visual Odometry Using Point and Line Features

Abstract

Robot localization with an RGB-D sensor has attracted substantial research attention lately. However, its robustness and accuracy are still poor in complex environments. To alleviate this issue, this paper presents a robust RGB-D visual odometry (VO) based on point and line features to allow accurate localization in complex scenarios including illumination change, low-texture and occlusion. Firstly, to enhance the robustness and accuracy of localization, we analyze the performance for different line feature extraction, and add the EDLine line feature to our VO system. Second, the EDLine lines are descripted using a sequence of point descriptors, and are matched by point correspondences and geometric constraints. Furthermore, a novel twin line reprojection errors optimization model is constructed by using the depth measurement information to correct the pose error. Finally, we integrate the line feature application into the RGB-D version VO part of ORB-SLAM2 to improve the estimation of the robot pose. The experimental results on public datasets show that our proposed VO approach outperforms other state-of-the-art VO solutions on accuracy and robustness under challenging environments.