Localization And Mapping Of Indoor Mobile Robots Based On RGB-D Cameras

Localization and mapping belongs to the underlying algorithm of the mobile robot platform.Because of the basis of other upper-layer applications such as logistics,floor cleaning and so on,it has extremely high research value and broad application prospects.This paper mainly focuses on the pose estimation of mobile robots when moving in an indoor environment similar to corridor structures.The innovation of this paper reflects in two aspects: first,we propose two plane detection methods,and second,under the constraint of the Manhattan World,based on the plane structures in the environment,a novel visual localization and mapping method is proposed.In this paper,we evaluate the proposed method extensively on real-world data and build the motion trajectory of the mobile robot as well as the 3D model of the environment where the robot is located.The experiment shows the good and robust performance of the proposed method.Our contributions focus on the following three points:1.This paper builds a robot motion and control platform carrying a few of related sensors and a computing device.The RGB-D camera has been calibrated for internal parameters,external parameters and depth offsets,and has obtained good depth measurement results as well as the accurate registration of RGB images and depth images.In this paper,a large number of experimental data are collected by the mobile robot equipped with an RGB-D camera,and some other parts of the data are collected from the published paper.For the subsequent work,all of the data are preprocessed.2.In order to make full use of the planar structure for robot pose estimation,we need to segment the plane in the scene.We propose two methods for plane segmentation: 1)The first method is based on the gradient change of the color,and the segmentation result of the color image is taken as prior information to accelerate the RANSAC route of the planar segmentation in a 3D point cloud;2)The second method uses only depth images,and proposes a fast planar segmentation algorithm based on disparity statistical histograms,which consists of horizontal and vertical histograms.By detecting line segments indirectly extracts the planar structure in the environment.The two plane segmentation methods proposed in this paper differ from other image segmentation in the fast speed.However the planar structure is not required to be segmented at the pixel level.The normal vector obtained by the approximate segmentation result can satisfy the subsequent robot pose estimation.3.This paper relaxes the constraints of the Manhattan world on the environment,and believes that the global indoor scene is composed of multiple local Manhattan worlds.At any position where the robot is located,we can construct one or more Manhattan coordinates based on the planar structures extracted from the camera.In order to eliminate the accumulated error as much as possible,we decouple the rotation and translation transformations.So we register the camera coordinate system to the dominant Manhattan coordinate system in the current state and realize the estimation of the camera rotation transformation,so as to eliminate rotational drift in the maximum extent.Based on the known rotation transformation,the translational motion is estimated by image feature matching or 3D feature matching according to the number of feature points that can be detected and matched in the images.