室内环境中基于点线的RGB-D视觉里程计

In this paper, a robust RGB-D visual odometry, based on point and line features, is proposed. It improves the localization accuracy of the visual odometry by using the geometric structure regularities in the indoor scene. This paper constructs a robust line extraction and matching method to improve the correctness of line matching. In this paper, the scene is modeled as a Mixture of Manhattan Frames, and multiple Manhattan systems in the scene are identified by line and surface normals. Different pose estimation methods are selected according to whether the Manhattan world hypothesis is satisfied in the scene: decoupled rotation and translation pose estimation method and based on feature-tracking pose estimation method. In this paper, the parallel and orthogonality constraints between lines and the constraints between lines and the dominant directions of Manhattan Frames are used to jointly optimize the camera\'s pose by combining the reprojection errors of points and lines. Experiments were carried out onsmall range of indoor scenessynthesized datasets and real-world datasets. The average localization accuracy in synthesized datasets is 1.5cm. In real-world datasets, average localization accuracy is 1.7 cm, which is twice higher than state-of-the-art methods. Experimental results show that the proposed method can effectively improve the localization accuracy and robustness of visual odometry in indoor scenes.