Monocular Visual SLAM with Robust and Efficient Line Features

Limin Long,Jinfu Yang

DOI: https://doi.org/10.1109/CCDC58219.2023.10326558

2023-01-01

Abstract:Point features are predominantly used in Visual Simultaneous Localization and Mapping(VSLAM). However, only using point features usually lacks the robustness in the scenes such as low texture, illumination variance, and structured environments. To overcome the above limitations, line features were actively employed in previous studies. But not all line features are helpful for improving the robustness of SLAM, since they are easily affected by image noise and feature detection strategies. Meanwhile, line features are easier to degenerate than point features due to parallel to epipolar lines. In this paper, we proposed a monocular visual SLAM using robust line features. First, a Robust Line Segment Detector (RLSD) algorithm is proposed to reduce the influence of noise and outliers by using gradient operators, length suppression, and line segment fusion. And a line feature matching method based on optical-flow is employed to reduce the instability in similar and structured environments. Then, a direction constraint strategy is proposed to solve the problem of triangulation failure due to degradation, which is conducive to subsequent mapping and tracking. Finally, a Visual-Inertial Odometry(VIO) with combined point-line features is implemented. Experimental results on the EuRoC dataset demonstrate that our proposed system outperforms other state-of-the-art methods.