Research On SLAM Based On Depth Perception Of Indoor Mobile Robot

Simultaneous Localization and Mapping(SLAM)is a method for estimating the pose of motion and constructing the map of the surrounding environment at the same time.It is a key issue in the research of mobile robot technology.The indoor SLAM based on depth vision can be divided into two-dimensional(2D)SLAM and three-dimensional(3D)SLAM with respect to the dimensionality of the map established.At present,the influence of 2D SLAM on depth image error is less,and the construction accuracy of environmental map is low.The3 D SLAM mostly uses point features in the structured low-textured environment,and the positioning accuracy is not sufficient.Based on the pseudo-laser data generation principle,this paper proposes a deep image inpainting algorithm with improved zero-point decision filtering.,and then a 2D SLAM system is built based on the improved pseudo-laser data to improve the accuracy of the environment map.At the same time,the feature detection method of point-line features is proposed in 3D SLAM,and the system construction is completed based on this,which improves the positioning accuracy.The main research contents are as follows:First,the depth vision based SLAM system architecture research.According to the SLAM system architecture,the characteristics of the depth camera are analyzed and compared,and the selection is established.The SLAM front-end compares the three ways of pose description and gives their range of application.The SLAM back-end analyzes the problem model of SLAM first,and then deduces the basic theory of nonlinear optimization.Finally,the construction of system architecture is completed based on the required system platform.Second,based on improved 2D vision SLAM research of pseudo-laser data.This paper calibrates the depth camera based on the plane calibration method,and analyzes the traditional pseudo-laser data generation method.For the cavity area of the depth image,the image inpainting domain is first determined,and then the improved algorithm of zero-point decision filtering is proposed.The improved pseudo-laser data is applied to the Gmapping algorithm to construct a 2D grid map of the environment.Again,a 3D visual SLAM study based on point-line features.In the SLAM front-end,the ORB point features in the image are extracted,and the feature number threshold is used as a condition for adding the improved LSD algorithm.The RANSAC algorithm is used to eliminate the error matching,and the reprojection error model and the analytical form of the Jacobian matrix are derived according to the matching relationship.In the SLAM back-end,the key frame selection method is established,and the closed-loop detection theory is fused to construct the pose-optimization model to obtain the global consistent trajectory and reconstruct the 3D map of the environment at the same time.Finally,the SLAM experiment research based on depth vision.In the real environment,the comparison experiments of the 2D SLAM system are carried out,and the map construction accuracy of the 2D SLAM system is verified by placing objects with different conditions.At the same time,ORB-SLAM2 is selected as the control of the 3D SLAM system.The multi-group comparison experiments under the ICL?NUIM dataset are used to verify the positioning accuracy of the system.Finally,the experimental research was carried out in the real environment to implement the map construction function.