Combined Point And Line Features RGBD SLAM Algorithm Research

With the rapid development of the society,the demand of mobile robot is increasing in both industrial and family fields.As one of the key technologies of mobile robotics,SLAM(Simultaneous Localization and Mapping)has a broad application prospect.And visual SLAM is an important research direction of SLAM,has achieved abundant research results.Although the point-based SLAM is effective in most cases,it fails to the camera tracking,or reduces its tracking accuracy in low-texture scenes or when the camera moves too fast.To solve this problem,in the thesis,PL-SLAM(Point and Line based SLAM),which is an extension of ORB SLAM and able to handle RGBD frames,is proposed to improve the tracking accuracy and reduce tracking failures.The main work includes:(1)The Line Features are added into SLAMFirstly,the extraction of the line features is studied.The Algorithm of EDLines(Edge Drawing Lines)is used to rapidly extract the keylines from the RGB images,gain the depth information of the keylines from the depth images and these keylines are divided in Mono keylines and Stereo keylines(the former can only apply to the matching of line features and the latter can also be triangulated into Map Line directly).The improved LBD(Line Band Descriptor)is used to describe the appearance information of line features and accelerate the speed for calculating appearance similarity between line features.Then,the matching of line features is studied.Appearance consistency check and geometric consistency check are used in the pixel coordinate system to determine whether the projected line segment matches the line features extracted from the images.(2)The line features are defined as Map Lines in 3D maps,and the parameterization of Map Lines and relevant applications of Map Lines in SLAM is studied.Firstly,the endpoint representation is used to parameterize the Map Lines.Due to the local observation of Map Lines,the method of projecting Map Line into line segment is given.Later,the re-projection error of Map Line is defined,so that SLAM can perform BA(Bundle Adjustment)with the observation information provided by the line features to optimize the camera pose in tracking,local mapping and loopclosing tread.And “Two-Step Optimization” process is proposed to eliminate the reprojection errors of the Map Line which might be caused during the optimization procedure.(3)A PL-SLAM combining Point and Line features is constructed.The PL-SLAM Algorithm is based on ORB SLAM to observe the point and line features and integrate the observation information into a back-end optimization model to optimize the pose as well as the map.It can be achieved through bringing the line features into the system and modifying the codes and the processing flows of the tracking thread,the local mapping thread and the loop-closing thread of ORB SLAM.And the modification of the three threads of ORB SLAM includes: in the tracking thread,the data preprocessing,the initial pose estimation,the local map tracking and new KeyFrame Decision are altered;in the local composition thread,the MapLine management module is added,the local map optimization and KeyFrame management are altered;in the loop-closing thread,the loop-closing detection and loop-closing correction are optimized by the visual dictionary and EPNPL(Efficient Perspective-nPoint and Line)Algorithm which are based on the combination of point and line features.The PL-SLAM algorithm is tested on the TUM RGBD benchmark.The results show that PL-SLAM algorithm does not only improve the robustness in the challenging environments,but also systematically improves the tracking accuracy in sequence frames.