Research And Implementation Of Vision-based Multi-robot Indoor Cooperative SLAM Algorithm

With the rapid development and demand of intelligent mobile robot technology,Simultaneous Localization and Mapping(SLAM)technology has been paid more and more attention.At present,single robot SLAM,especially the single robot SLAM based on visual camera,has made great progress in the academic field and application.But the main research direction of the current researchers is still more concentrated on single robot SLAM.The research on multi robots cooperative SLAM is relatively rare.According to the demand of multi robots cooperative auxiliary operation in future factory,this paper make a preliminary research and experiment on the multi robots cooperative SLAM algorithm based on vision.This paper firstly studies the basis of single robot algorithm in multi robots cooperative SLAM,introduces the mathematical description of the SLAM problem and the selection of the sensor,and makes a survey and summary of the front and back theory of the SLAM algorithm.After the comprehensive comparison of the requirements,the ORB-SLAM algorithm is selected as the foundation of the implementation,and the implementation framework of the algorithm is also introduced.The main research hotspots of multi robot cooperative SLAM include three aspects.The first is task allocation and multi robots communication.The second is data association of multi robots(or cameras)position.The third is map splicing depending on data association(some researchers add data association into map stitching).In this paper,the research of multi robots algorithm mainly focuses on the map splicing,especially on the analysis of two cases of map splicing.The first is a map mosaic algorithm based on the relative observation of the cameras.In this case,the ICP algorithm is used to solve the phase position and to optimize it by BA in this paper.High precision relative transform matrix could be gotten,and then coordinate transformation of map point cloud is carried out to realize map splicing.The second situation is map mosaic algorithm based on scene identification.In this case,the robots do not meet each other but are passing through similar scene.At this time,it needs scene recognition to judge whether the map can be fused.The visual bag of word technology is used to identify the scene.By importing the image dictionary tree,the reserved image frames are classified and numbered according to the dictionary word information,and the same image index is used to identify the scene at any time,and the similar image is successfully matched.After matching the similar image frames successfully,we use the Pn P algorithm and the BA optimization process to solve the multi robots relative position and pose transformation relationship in this scene,and convert the local map which it has constructed to the coordinate system in this scene,and then the map splicing is carried out according to the relative position and attitude relationship.The specific algorithm implementation process framework is given.The experimental part of this paper includes the data set test of the single robot algorithm,which proves that the tracking performance of the ORB-SLAM algorithm is superior.Through the experimental test of the algorithm,the sparse point cloud map is constructed.Because of its sparsity,the algorithm has the feasibility of carrying out the real-time map splicing,and the experiment is implemented for the multi robots algorithm.For implementation experiment of multi robots algorithm,a simple and feasible communication scheme is implemented based on ROS system.One robot is used as a server and a client.The other is only the client and they connect through the Wi Fi.The client releases its own observation and state information at any time through the topic and releases its own built environment map when necessary.The purpose of this experiment is to verify the effect of multipoint cloud map splicing,because the sparsity of the sparse point cloud is relative,so the experiment is mainly to make map splicing in a small range and evaluate it,which is the basis of large scale expansion,and the experimental results show that the stitching effect is great and could meet the expected goals.