Research On Monocular Vision SLAM Of Indoor Inspection Robot Based On Point And Line Feature

SLAM is a real-time localization and mapping technique widely used in fields such as robotics,autonomous driving and AR/VR,and visual SLAM algorithms represented by feature-based methods have been extensively studied due to their good stability and computational efficiency.At present,point feature visual SLAM has the problem of inaccurate localization or localization failure in low-texture environments and camera rapid movement scenes.This paper proposes a point-line feature fusion SLAM system to improve the shortcomings of using single feature methods.Firstly,the related theories of Visual SLAM were introduced,including the transformation from the camera and IMU coordinate systems to the real coordinate systems,and the analysis of the errors caused by the sensors,as well as the calibration experiments to obtain the related parameters and establish the transformation relationship.The knowledge of line feature parameterization and initialization was also introduced to provide the theoretical basis for camera pose estimation.Secondly,the point-and-line fusion method was used in the front-end visual odometry to improve the localization accuracy of the inspection robot.To solve the problem of large number of short lines and similar line segments in the feature point detection process,the short lines were eliminated and the similar line segments were merged.The motion posture of the camera was estimated after the feature points of the adjacent images were matched,and the geometric constraint method was used to realize the line feature tracking of adjacent frames,which saves the time to calculate the LBD descriptor and improves the matching efficiency compared with the method of using LBD descriptor for matching.Then,a point-and-line fusion VI-SLAM system was set up,and the line segment detection method and matching method proposed in this paper were applied to the system.To solve the scale uncertainty problem of the monocular camera,the visual and inertial information were fused to establish a visual-inertial tight-coupled optimization model.After the visual initialization,the initial posture of the camera was obtained,and then the sensor observation data was aligned to restore the real scale.The accuracy of the trajectory was improved through the loop detection function.Finally,The proposed method reduces the absolute root mean square error of the trajectory on the TUM-VI room sequence and the difficult sequence of Eu Roc dataset by 38% and 20%,respectively,compared with the VINS-Mono method.In addition,the test of the indoor inspection robot in the actual scenes such as corridor,corridor and hall also proves that the point-line feature fusion algorithm has higher robustness.