Research On Simultaneous Localization And Mapping Technology Of Indoor Mobile Robot Based On Visual And IMU Fusion

With the diversification of the application scenarios of indoor mobile robots,people demand higher positioning accuracy of mobile robots.Simultaneous Localization and Mapping(SLAM)technology is one of the effective ways to solve the positioning problem of mobile robots indoors,which has become a hot topic of autonomous unmanned system research field.Since the SLAM technology based on the visual sensor is not yet mature,there are still many problems in the environment such as lack of texture,and considering there is a principle that the Inertial Measurement Unit(IMU)and camera have certain complementarity,this paper conducts an in-depth study on the simultaneous localization and mapping technology for indoor mobile robot based on vision and IMU fusion,which has certain significance for accelerating the large-scale application of indoor mobile robots.The main research contents of this paper are as follows:The overall design of the binocular visual-inertial SLAM system is carried out in view of the current demands of deploying visual-inertial SLAM of indoor mobile robots to avoid scale drift,enhance the accuracy of loop detection,and realize map navigation.Build an indoor mobile robot test platform,and carry out the overall design of the indoor mobile robot from the mechanical structure,hardware system,and software system.Establish the world coordinate system,the mobile robot coordinate system,the sensor coordinate system and determine the conversion relationship between the coordinate systems,and the data model of the binocular camera and the IMU are established,and realize the calibration of the internal and external parameters of the sensors.The test results show that the the reprojection error of camera and IMU calibration is less than 0.5 pixels.Aiming at the problem that the uniformity of feature point extraction and feature point matching accuracy affect the accuracy of front-end odometry pose calculation,a quadtree structure is adopted to homogenize the feature point extraction;the sub-pixel level matching of binocular feature points is realized based on the SAD algorithm,which improves the accuracy of feature point depth calculation;to reduce the probability of feature point mis-matching,a mis-matched feature point elimination strategy combining feature main direction and RANSAC is proposed;the test results show that compared with the traditional ORB feature extraction algorithm,the feature extraction method in this paper increased the extraction speed and image uniformity by 9.66%and 45.29%,respectively;compared with the FLANN algorithm,the matching accuracy of the feature point matching algorithm in this paper has increased by 16.23%.A loosely coupled approach is used to initialize the variables of the visual-inertial SLAM system,analyze the visual constraints,IMU measurement constraints,and prior constraint principles for back-end optimization,and reduce the front-end odometry cumulative error based on the sliding window strategy by using graph optimization to jointly optimize the state quantities such as pose and IMU bias.Aiming at the problem of unstable loop detection when the indoor environment lacks significant landmarks,a loop detection algorithm based on the fusion of AprilTag and bag-of-words models is proposed to improve the accuracy and precision of loop detection.Aiming at the problem that sparse point cloud maps are difficult to be used for mobile robot navigation,a method of splicing dense point cloud based on key-frame poses to generate octree maps is proposed.The actual scene mapping test shows that the method proposed in this paper reduces the size of point cloud by 98.09%,and the map built by the method can describe the characteristics of the test environment.Compare the VI-SLAM algorithm in this paper with VINS-Mono and ORB-SLAM2(binocular mode)to verify the positioning accuracy and real-time performance of the system.Tests are carried out on four representative datasets of EuRoC and the actual indoor environment.The results show that the positioning accuracy of the proposed algorithm is 17.70%and 28.33%higher than that of VINS-Mono and ORB-SLAM2,respectively,on four representative datasets in EuRoC.In the actual scene,the algorithm in this paper has better trajectory consistency,more accurate positioning,and correct loop detection ability.The root mean square error of its absolute trajectory is 0.749m,and the absolute trajectory length error is 0.238m.The operating efficiency of the VI-SLAM algorithm in this paper is tested.The average processing time of a single frame in the front-end odometry in this paper is 127.056ms,which is 18.60%faster than ORB-SLAM2.