Research On RGB-D SLAM System Fusing Monocular Information

Simultaneous localization and mapping(SLAM)is a key technology to help robots realize autonomous positioning and navigation.SLAM typically uses sensors such as lidar,camera,inertial navigation module and code disc to capture external information.With the advent and development of RGB-D cameras,they are gradually applied to the fields of positioning,mapping,and ubiquitous computing,and the research on RGB-D SLAM has also become one of the important topics in the field of SLAM.RGB-D camera can provide rich color and depth information at a relatively low cost,but there are also problems such as limited field of view,narrow measurement range and large noise when using it.Many SLAM systems based on RGB-D camera rely heavily on depth information and are limited by the type and geometry of the environment.Therefore,this paper has done a thorough research on this and proposed an RGB-D SLAM system fusing monocular information based on the existing RGB-D SLAM system by making full use of the information of RGB and depth images.The specific research contents are as follows:Considering the field of view and measurable distance of camera,this paper proposes a depth selection mechanism to estimate the depth image from the existence,accuracy and discreteness of depth information in the front end of slam.If the depth image meets the three critical conditions proposed by the selection mechanism,which can be considered to have sufficient depth information,and the system will start SLAM work with a combination of color image and depth image;Else the depth information is considered insufficient and the system will only use color images to SLAM work.The system is divided into RGB-D and MONO through a selection mechanism.Because the system estimates the initial pose of the camera with motion model or reference frame,the difference between the two processing methods of RGB-D and MONO is specifically reflected in the optimization stage.The system uses BA optimization algorithm to optimize the initial pose based on the information provided by RGBD or MONO.For the RGB-D method,the pose is optimized based on the three-dimensional information(including depth information,etc.)provided by the image.As for the monocular method,the pose is optimized based on the two-dimensional information provided by the separate color image.The system also includes loop detection and global optimization modules to further optimize the pose,meanwhile it establishes a dense three-dimensional point cloud map based on the information provided by the key frames.Finally,the feasibility,accuracy and robustness of the system are verified by experiments.