Map Construction Of Mobile Manipulator Navigation Based On Multi-sensor Fusion

Real-time map construction is the online visual presentation of the surrounding environment perception of the mobile manipulators,and it is also an important basis for the mobile manipulators to complete the navigation task.The design structure of the map affects the efficiency of the mobile manipulators in completing the operation instructions in the navigation task.With the development of sensors and the improvement of computer processing performance,mobile manipulators navigation has gradually entered the stage of multi-sensor fusion from the single-sensor stage,and the information source of map construction has also changed from single-modal information to multi-modal collection.The map construction methods based on multi-sensor fusion can give play to the advantages of different sensors and improve the accuracy and speed of map construction,so it has important research significance.At present,the errors and noises caused by the construction of maps based on a single sensor can not be avoided;the construction of maps based on 3D voxels is slower and the files take up too much space;the methods based on pure visual loop detection are less robust and has a higher false detection rate.In view of the above three problems in the construction of the mobile manipulators navigation map,the paper proposes three methods based on multi-sensor fusion to improve:(1)Aiming at the noise generated by 2D Li DAR and RGB-D camera used in map construction,a fusion denoising method for 2D Li DAR data and depth images alignment is proposed.This method extracts multiple sets of 1D sequences from depth images and 2D Li DAR data,aligns the sequence groups by calculating the similarity of the point sets,and then uses random sample consistency and homography to determine the noise area,and then repairs noises by complementary fusion between aligned sequence groups.The data processed by this method can be used to construct 2D maps and improve the visualization of the depth images,and effectively supplement the map information missing from the original data of a single sensor.(2)Aiming at the limitation of building a map with a single sensor,a multi-sensor fusion map building method based on current location storage is proposed.This method designs a map construction method that stores all multi-modal information in the current grid.The multi-frame 2D Li DAR data and the mobile manipulator's pose are optimized by the iterative nearest point algorithm,and the image data is convolutional auto-encoding to reduce the occupied space,and all-modal data are associated and stored by a graph structure.Compared with the independent 2D Li DAR mapping method,the proposed method adds the RGB-D visualization function.Experiments have proved that this mapping method is about 85%faster than RGB-D point cloud mapping under the same conditions,and the sizes of the map files are about 30% smaller than that of RGB-D point cloud mapping.(3)Aiming at the limitations of the visual closed loop detection methods,a closed loop detection method combining appearance and pose information is proposed.This method first uses the convolutional autoencoder and the k-nearest neighbor algorithm to filter out the scene images with high similarity,and then uses the siamese network to further determine whether the scene is the same;on this basis,the mobile manipulator pose information is introduced in at the decision-making level.The authenticity of the closed loop is verified through the muti-sensor information fusion.After experimental verification,the appearance discrimination method in this method is similar to the traditional classical method in accuracy,but the average speed is increased by about 63%;the average accuracy of the overall closed loop detection method is 89.97%,which is about 6? 7 percentage points higher than the classical method.