Research On 3D SLAM System Based On Panoramic Vision

Simultaneous Localization and Mapping(SLAM)has always been the core subject in robot research.As the basic function of the robot,the cognitive level of environment directly influences the success rate and complexity of robot task.SLAM research based on visual sensor has become the mainstream of SLAM research because of its low price.In recent years,panoramic camera has gradually entered the market with the development of Virtual Reality(VR)technology.Compared to normal camera,it can capture all visual information in all directions at once.It gives the visual SLAM system lot more room for performance improvement.The research of panoramic visual SLAM system here is based on panoramic camera.Panoramic camera is usually composed of several wide-angle lenses.The panoramic camera used in the experiment consist of two fisheye lenses mounted back to back.First,the calibration of fisheye lens is studied.The radial distortion correction of fisheye lens based on cylindrical checkerboard template is presented.The spherical image is projected into a plane image and a panoramic image according to the mapping relation.The experimental results show that the method is effective for correcting the radial distortion.For SLAM system,the fisheye images is projected into plane image according to the pinhole camera model.Pixels are projected back into the camera coordinate system.To satisfy the mapping of large scene,the direct method is used as the core of the panoramic visual SLAM system.And the state estimation is based on the nonlinear optimization method.Instead of using feature point matching,the method of tracking map points is adopted.Map points are described by keyframes and inverse depth,which reduce the variable dimension in the optimization process.In addition,the robustness of direct method is improved by adding photometric affine transformation equation.The nonlinear optimization equation is created based on sliding window method.The map points are projected to the keyframes managed by the sliding window for parameter optimization.A sparse point cloud map is created by keyframes and pose information of two fisheye cameras.The experimental environment is selected in the indoor scene.The experimental data is collected by a vehicle,and used in the panoramic visual SLAM system described in this paper.The visual experiment results show that the panoramic view can help environment reconstruction and localization completed more effectively.Experimental data analysis shows the reliability of the system described in this paper.