Research On Monocular VINS System Based On Dynamic Fusion Of Point-line Feature

The application of mobile robots to replace manpower in uncharted fields has attracted increasingly interest,and Simultaneous Localization and Mapping(SLAM)is the indispensable technology.Most current visual SLAM technologies use point features in the environment as tracking objects,but once in a weak texture environment,this method will cause large positioning errors or even failures due to the lack of sufficient feature points in the weak texture environment.However,the line feature in the environment can be used as a supplement when the point feature is insufficient to constrain the motion pose of the robot.Therefore,based on the monocular vision camera and Inertial Measurement Unit(IMU),this thesis studies and designs a monocular visualinertial system(VINS)for dynamic fusion of point and line features.The main research contents and methods of this thesis include:Firstly,according to the point and line features of the surrounding environment,the ShiTomasi corner algorithm and improved LSD(Line Segment Detector)algorithm are used to extract the point and line features on the image,and the KLT(Kanade-Lucas-Tomasi)optical flow algorithm and the LBD(Line Band Descriptor)descriptor are used to realize the matching and tracking of the point and line features.In order to facilitate the line feature to participate in the optimization of nonlinear pose estimation,the Plücker coordinate and orthogonal representation are applied to the representation of spatial line features.At the same time,the reprojection errors of point features and line features are constructed for nonlinear optimization in the backend.For IMU measurement,the pre-integrated measurement value is used as the IMU measurement error for back-end nonlinear optimization.By constructing the quantitative relationship between point and line features to filter most useless short lines,formulating a short line elimination strategy,and adjusting the two hidden parameters of the LSD algorithm,the improvement of the traditional LSD line feature extraction algorithm optimizes the quality and efficiency of line feature extraction.Then,the sliding window model is adopted in the state estimation of the back-end to keep the number of keyframes constant and reduce the computational dimension.Two marginalization strategies are used to optimize the variables in the window,and the constraint information removed from the sliding window after marginalization is converted into prior information.The prior information,IMU measurement error,the reprojection errors of point features and line features and loop closure detection error are jointly constructed into an overall error function to estimate the pose.Among them,the loop closure detection error only participates in the optimization when a closed loop occurs.Secondly,in order to avoid numerous accumulated errors caused by the long-term operation of the system,the bag of words model is used for loop closure detection.When the system detects that a closed loop occurs,the sequential edge and loop-closure edge are added to the pose graph for pose optimization,and the pose is corrected in real time.The global trajectory consistency is achieved,which ensures the stability of the system for a long time and a wide range of operation.Finally,on the EuRoC dataset,the experimental results and accuracy of this system algorithm are compared with VINS-Fusion(stereo without loop),VINS-Mono and PL-VINS.The experiments show that the proposed algorithm not only improves the speed of line feature extraction,but also provides consistent local and global pose estimation.The algorithm proposed in this thesis is engineered on the camera of Mynteye,building the Bulldog-CX robot experimental platform,and two challenging indoor and outdoor real scenes are selected for experiments.The experimental results show that the monocular VINS system with dynamic fusion of point and line features designed in this thesis has strong robustness in a wide range of indoor and outdoor scenes.