Point-line-based RGB-D SLAM System In Indoor Environment

At present,visual Simultaneous Localization and Mapping(SLAM)technology is applied in different fields,such as automatic driving,AR,VR,warehouse robots,etc.However,the visual SLAM in the indoor environment is often affected by factors like low texture,repetitive structure,and illumination changes,which makes the SLAM system based on point features unable to detect rich feature points,and is difficult to perform feature matching and tracking normally in camera ego-motion estimation,resulting in the reduction of the localization accuracy and robustness.Based on the above problems,using an RGB-D camera as the sensor input,this paper proposes a point-line-based visual SLAM system in the indoor environment,which uses the structural regularities in the scene to enhance the localization accuracy and robustness in low-texture and repetitive structure scenes.At the same time,the construction of dense maps will enrich the environmental information,which is conducive to the subsequent path planning and navigation tasks.The detail content is as follows:(1)Aiming at the problem of low line feature matching accuracy and many mismatches,a new line feature mismatch removal method is proposed,which uses the overlapping degree and geometric relationship of line features to remove line feature mismatches and improve the accuracy of feature matching.(2)For the robustness of pose estimation in different indoor scenes,a different pose estimation strategy is selected according to whether the detection scene is a Manhattan World(MW)scene.Among them,for the problem of scene recognition and matching,it is proposed to use the space line and surface normal to identify the Manhattan Frame(MF)existing in the scene,and then classify the line as a structural line,and use the structural line to match the Manhattan Frame.Different camera pose estimation methods are selected based on the Manhattan Frame recognition and matching results to improve the accuracy and robustness of visual odometry pose estimation.(3)Aiming at the problem of further utilization of structural rules in the scene,a local map optimization method based on point-line features and scene structural constraints is proposed.The reprojection error of points,the reprojection error of lines are combined with the parallel and orthogonal constraints of lines,and the constraints between structure lines and Manhattan systems are jointly constructed to optimize the camera pose and feature position.Finally,on the basis of the ORB-SLAM2 system,the above functions are built,and the dense mapping function based on surfels is added,so as to realize the RGB-D SLAM system based on point-line features in a complete indoor scene.Extensive experiments on ICL_NUIM datasets,TUM RGB-D datasets,and actual scene experiments are conducted to compare the performance of the proposed method with similar methods.Results show that the proposed system can achieve higher localization accuracy and robustness compared to other systems.At the same time,the ablation experiment is constructed to verify the effectiveness of the different modules proposed in this article.