Development Of Visual SLAM System Based On Panoramic Camera

Simultaneous Localization and Mapping(SLAM)is recognized as the frontier direction of space positioning technology in the field and has achieved remarkable results in the past 30 years.It mainly solves the problem of positioning and mapping of the mobile robots in an unknown environment,which mainly used in robot autonomous investigation,inspection and intelligent transportation.Visual SLAM has become a research hotspot in this field because of its low price,large amount of information and ease of installation.Pinhole cameras are the most widely used cameras in the field of machine vision.However,due to the limited field of view of pinhole cameras,there are deficiencies in the stabilities and accuracy in camera pose estimation.Compared with the pinhole cameras,the panoramic cameras have a 360° field of view,which means that the camera can simultaneously acquire all the features in the environment,maximizing the amount of information obtained by the camera.And provide camera estimation and mapping more sufficient data.In this paper,a visual SLAM algorithm based panoramic cameras is proposed,which improves the accuracy of the real-time camera pose estimation by using the advantage panoramic cameras' large field of view,and build the sparse feature map at the same time.Experiments show that this algorithm is suitable for indoor and outdoor environments and provides an effective solution for mobile robot localization and navigation tasks.The main research work include:1.The spherical model of the panoramic camera is established,and the method of system initialization is improved according to this model,which improves the success rate and accuracy of initialization.In the initialization process,the camera poses are calculated using the features of specific regions.This avoids the erroneous results of the essential matrix decomposition.And then the initialization results are optimized,which improves the initialization success rate and improves the accuracy of the camera.2.Redefine the reprojection error of the camera for the panoramic camera spherical model,and derive the Jacobian based on this reprojection error in bundle adjustment and pose optimization.Finally,use the least square method to iterate the best estimate of the camera pose and map points.3.Design a complete visual SLAM system that includes: system initialization,visual tracking,map construction,closed-loop detection,and re-localization.The visual tracking,map construction and closed-loop detection are performed in three independent threads respectively,which ensures the real-time performance and function implementation at the same time.