Research On Robust Visual SLAM Based On Multiple RGB-D Cameras

With the development of science and technology,robots have been integrated into all aspects of people's production and life.As the key technology of mobile robots.Simultaneous Localization and Mapping,(SLAM)is now a research hotspot in the robot field.Because of its low cost and rich visual information,visual SLAM technology has attracted wide attention in SLAM field,and is an important research direction of 3D environment mapping.Among all visual SLAMs,RGB-D camera-based vision SLAM is a hot topic in recent years.Due to the narrow field of view(FOV)of single camera.it is greatly affected by noise and local illumination,texture features and viewing angle,leading to the reduction of localization accuracy and even the failure of tracking during the operation of SLAM.In this dissertation,multiple RGB-D cameras are used to increase the FOV,in order to improve the accuracy and robustness of visual SLAM in complex environment or environment lacking visual features.The main work of this dissertation is as follows:First of all,in order to find out the relative poses between cameras with non-overlapping FOV,a two-step extrinsic calibration method is proposed,which is used to calibrate the multi-camera system(MCS)with non-overlapping FOV configured in any direction.Firstly,a sparse map of the environment is built to estimate the initial calibration of the MCS through PNP for each camera,and then a pose graph is constructed to obtain an accurate extrinsic calibration by optimizing the pose graph with the initial estimate.Secondly,aiming at the front-end problem of visual SLAM with multiple RGB-D cameras,a visual SLAM system model based on multiple RGB-D cameras is proposed,including the multi-camera model and the map representation.In order to deal with the complex mapping in the multi-camera system,an MCS coordinate system is established in the multi-camera model as the intermediate coordinate system when mapping points between the camera coordinate system and the world coordinate sy stem.In map representation,multi-keyframe is introduced to store images from different cameras and corresponding features,and to fuse map points created by different cameras.Through the multi-camera model and multi-keyframe in the proposed multi-camera SLAM,it is not necessary to calculate the motion of each camera separately,and the input of all cameras will be used to estimate the pose of the multi-camera system directly.Because of the complex mapping in the multi-camera system,an MCS coordinate system is used as the intermediate coordinate system for map point mapping between the cameras coordinate system and the world coordinate system in the multi camera model.Therefore,it is not necessary to calculate the motion of all cameras separately,and the input of all cameras will be used to estimate the pose of the multi-camera system directly.In map representation,multi-keyframe is introduced to store images from different cameras and corresponding features,and to fuse map points created by multiple cameras.Then,in the back-end of the multi-camera visual SLAM method,in order to build the optimization model,a correct objective function is needed.In this paper,a graph optimization method based on multi-camera model is adopted,where the reprojection error with multi-camera system is defined by utilizing the constraints between cameras,and the Jacobian matrixes in bundle adjustment and pose graph optimization are derived respectively.Thus,the optimization model is constructed.Finally,a visual SLAM system is built,and experiments are designed to verify the effectiveness of the proposed visual SLAM method.Experiments are carried out on the condition that the multi-camera system is mounted on the mobile platform and is held by hand.The experimental results show that the visual SLAM based on multiple RGB-D cameras has high accuracy and robustness both in plane motion and 3D irregular motion.