Line-based visual odometry using local gradient fitting

Visual odometry aims to estimate the relative pose between frames, which is a fundamental task for visual SLAM. In this paper, we present a novel line-based visual odometry (VO) algorithm that fully utilizes the characteristic of line to estimate the projected line of adjacent frame by minimizing the local gradient fitness evaluation. In contrast to the current feature-based or line-based visual odometry, we don′t need to explicitly match points or lines of two frames, which is non-trivial and inaccurate in challenging scenarios such as texture-less scenes. In our method, the projected line is calculated simultaneously with the local gradient fitting function of pose estimation based on the constraint that the orientation of the projected line should be perpendicular to the gradient orientation of pixels of its local regions. The proposed method is more robust and reliable than other line-based VO since it fully uses the pixel orientations in the local regions to estimate the projected line and relative pose. We evaluate our method on the real-world RGB-D dataset and synthetic benchmark dataset. Experimental results show that our method achieves the state-of-the-art algorithms in indoors scenes, especially in texture-less scenes. •A visual odometry based on line features and RGB-D camera.•The local gradient fitting of line is used to construct the error function.•A keyframe selector suitable for texture-less scenes is proposed.•The robustness of visual odometry in texture-less environments is improved.