Research On Localization Of Indoor Mobile Robot And Reconstruction Of Semantic Point Cloud

Simultaneous localization and mapping technology is one of the key technologies of indoor intelligent mobile robots.Integrated navigation positioning and semantic reconstruction technology based on visual inertial data fusion is the current focus of research in the field of intelligent mobile robots.To meet the mobile robot localization and perception of the environment and other needs of the current positioning error drift over time for inertial navigation systems,visual navigation system vulnerable to environmental factors,the traditional three-dimensional point cloud reconstruction lack of understanding of the scene of a series of problems Carry out research on the tight combination positioning of visual and inertial data and semantic reconstruction technology,aiming to solve the problem of indoor intelligent mobile robot positioning and 3D point cloud reconstruction.This paper uses binocular camera combined with inertial measurement unit to complete the robot positioning and semantic point cloud reconstruction algorithm in indoor scenes.The main work of this paper is as follows:(1)Aiming at the positioning problem of indoor intelligent mobile robots,this paper uses a combined navigation positioning algorithm based on the combination of binocular cameras and inertial navigation,and gives a filtering fusion positioning algorithm based on extended Kalman filter and an optimized fusion based on factor graph optimization Location Algorithm.(2)Aiming at the classic visual-inertial integrated navigation algorithm based on the extended Kalman filter,the numerical stability of the algorithm is analyzed,and an IMU data preprocessing algorithm based on the closed solution pre-integration of the IMU is proposed.The algorithm pre-integrates the IMU Solving a closed solution to replace the traditional median integral solution can effectively improve the accuracy of the algorithm.This paper also proposes a camera observation equation representation method based on inverse depth parameterization.This algorithm can effectively solve the problem that exists when the threedimensional space point z-axis in the camera coordinate system approaches 0 by parameterizing the three-dimensional point coordinates in the camera observation equation.The singularity problem improves the numerical stability of the algorithm,and the data set simulation experiment verifies the stability and performance advantages of the algorithm compared to the classic algorithm.(3)Aiming at the traditional visual-inertial integrated navigation algorithm based on nonlinear optimization,the factor graph theory and the derivation of the optimization function are described.The data set experiment verifies the advantage of the factor graph optimization algorithm in positioning accuracy compared with the traditional optimization algorithm,and through the calculation algorithm The CPU consumption and memory consumption of this test verify the advantages of the factor graph optimization algorithm in performance and system consumption.(4)Aiming at the 3D reconstruction algorithm of the indoor scene semantic point cloud,this paper combines the Mask R-CNN semantic segmentation network and the point cloud voxel clustering algorithm to reconstruct the 3D point cloud of the scene.The algorithm first constructs a three-dimensional point cloud of the scene through the depth image,RGB image and the pose of the camera,and at the same time performs semantic segmentation on the RGB image to obtain the semantic label of the image,and then the three-dimensional point cloud is clustered on the point cloud by supervoxel clustering.Perform segmentation,and finally map the two-dimensional semantic tags to the three-dimensional point cloud to obtain a globally consistent semantic point cloud reconstruction.(5)An intelligent mobile robot prototype experimental platform was built,and the visual inertial combined positioning algorithm based on IMU pre-integration closed solution and the visual inertial combined positioning algorithm based on factor graph optimization were realized through the prototype software and hardware.Based on the prototype platform,a number of positioning experiments are designed to verify the effectiveness and effects of the algorithm in this paper,including indoor linear motion experiments,absolute trajectory error positioning experiments,curve experiments,and long-distance long-distance positioning experiments.And through the physical camera recording data,through two different real scenes to verify the effectiveness of the semantic point cloud reconstruction algorithm in this article.