Robust Monocular Visual Odometry with Point-Line Features in Low-texture Scenes
Hao Wu, Haiyang Yu, Xingru Qu, Rubo Zhang · 2023
Traditional visual SLAM systems rely on tracking point features for camera pose estimation and environment map construction. However, these tasks remain challenging in low-texture scenes where the number of available features is limited. This paper proposes a novel visual odometry solution that incorporates both point and line features. To be robust in low-texture scenes, the proposed method develops the LSD algorithm to detect line features, which includes short lines rejection and line segments merging to improve the identification of line features. Validating our approach, we conduct experiments on the EuRoC and TUM datasets and compared our results with ORB-SLAM2 and PL-SLAM, which demonstrate that our method not only improves the quality of extracted line segments but also reduces the time consumption required for line segment tracking. Additionally, the visual odometry presented in this paper shows greater robustness in low-textuer environments.