Research On Large-Scale Map Management And Loop Detection Method Of Visual SLAM

With the deepening research of artificial intelligence technology,robotic autonomous navigation technology has entered a stage of rapid development.Mobile robots analyze the data acquired by laser or vision sensors in unknown environments to create maps,and then guiding robot autonomous positioning and navigation.Simultaneous Localization And Mapping(SLAM)is one of the most popular robot positioning and navigation technologies.It uses feature matching for environmental map construction and robot positioning.Consider the importance of managing large-scale maps and optimizing robotic pose through loop closing when robots work in complex scenarios and require long hours of operations.This paper mainly studies some key problems of map management and loop detection in visual SLAM,including the research of large-scale map management and optimization methods in vision SLAM,the research of map storage methods and the research of loop detection methods.This paper proposes some improvements to solve the key problems in visual SLAM.The main work of this paper is as follows:1.Research on large-scale map segmentation method based on graph theory.For the large and complex map,this paper implements the segmentation method of the SLAM map based on graph theory.We divide a large-scale,complex map into several sub-maps.At the same time in the working mode,we save memory by loading the partial map,and use the divided partial map instead of the whole map for global optimization.2.Optimization of loop detection based on bags of visual words.This paper will study the loop detection method of visual SLAM.The method of combining L1 norm and chi-square distance is proposed in this paper to calculate the similarity score between images for improving the accuracy of loop detection in visual SLAM.3.Design of online map generation and offline map loading algorithm in visual SLAM.For some mobile robots who work in a certain scene for a long time,this paper studies how to generate,save and load SLAM maps.The method proposed in this paper divides the SLAM system into learning and localization modes.We save a detail map as an offline map in learning mode and use it to realize the localization of mobile robot in working mode.4.Experiment on the reliability and accuracy of large-scale map management and loop detection.The validity and the location accuracy of the proposed algorithm are validated by using RGBD camera to capture the real time images and standard data sets.