Study On Visual Localization Algorithm Of Mobile Robot In GIS Scene 3D Model

As the key technology of mobile robots,real-time positioning and mapping have extensive application value in intelligent manufacturing,battlefield reconstruction and other fields.However,the existing methods based on continuous frames gradually increase the accumulation of errors with the long-term operation of the mobile robot,which leads to the phenomenon of pose drift and large positioning error of the mobile robot under long-term work.In view of this,this article starts with the causes of errors in the existing real-time positioning and mapping.Aiming at the causes of the problem,it proposes an algorithm for the global positioning of mobile robots in the tilted aerial photography model,and analyzes the problem of using the tilted aerial photography model as a positioning map.Where and solve it,and finally realize the global positioning of the mobile robot in the model.First of all,this article explains the basic principles of real-time positioning and mapping.By constructing a real-time positioning and mapping system,experiments are conducted to verify the influence of the accumulation of errors on the positioning accuracy of the existing incremental method,and analyze the causes of errors.It is: the map is incrementally built by calculating the change of the pose through the feature matching between consecutive frames,and the error of each frame will accumulate frame by frame,and it is difficult to completely eliminate it even through back-end optimization.In view of this,a method of global positioning in the oblique aerial photography model is proposed,so as to avoid the accumulation of errors based on incremental real-time positioning.Secondly,build a tilted aerial photography model based on the proposed scheme and analyze the shortcomings of this model as a mobile robot positioning map: tilted aerial photography lacks data at blind spots of the viewing angle,resulting in many defects compared with the real scene,which affects the use of the robot model.As the positioning accuracy of the positioning map.At the same time,a solution is proposed:by collecting point cloud data at the blind spot of the oblique photographic perspective on the ground,and registering it with the oblique aerial photographic point cloud data,to supplement the lack of point cloud data at the blind spot of the viewing angle,thereby eliminating the defects of the model Repair it.At the same time,research on ground point cloud denoising algorithm and registration method of air and ground point cloud data are carried out on the point cloud data collected on the ground.Finally,using the repaired oblique aerial photography model as the positioning map,the simulated camera collects images and the corresponding camera pose in the model to make a data set,builds a pose estimation network based on deep learning for training,and finally realizes the mobile robot in the model In the global positioning,the positioning accuracy of position and angle are 3.85 m and 7.73°,respectively,which is the best accuracy compared to directly adopting the unrepaired model for positioning:20.16 m and 16.98° are greatly improved,and in the actual mobile robot Deploy on the mobile robot to realize the practical application of global positioning of mobile robots.It avoids the problems of pose drift and positioning error accumulation caused by SLAM’s incremental construction of maps for positioning and navigation,and broadens the practical application scenarios of mobile robots.