Study On Tracking And Positioning Of Visual-inertial Odometry

Visual Simultaneous Localization and Mapping(VSLAM)has achieved remarkable development in applications such as robotic autonomous navigation,unmanned driving,virtual reality and augmented reality.However,in environments with rapid rotat ion or lack of texture,VSLAM relies on a single visual sensor,resulting in poor positioning accuracy.In order to enable the VSLAM positioning algorithm to cope with complex and changing environments,multi-sensor information fusion has become the focus of research.An inertial measurement unit(IMU)consisting of an accelerometer and a gyroscope can measure the angular velocity and acceleration information of an object,and the sampling frequency is high,which can adapt to rapid movement.The fusion of IMU and VSLAM can effectively improve the positioning accuracy,but there is a huge challenge in multi-sensor information fusion at different frequencies.Therefore,this paper mainly conducts in-depth theoretical research and exploration of VSLAM's multi-sensor fusion algorithm of vision and inertial measurement,and considers the real-time and robustness problems of VSLAM algorithm.The main research contents and results are as follows :(1)To address the problem of time-consuming in feature matching of the featurebased SLAM,this paper take advantage of the rapidity of optical flow tracking,uses ORB features and improved optical flow in feature matching to improve the real-time performance of the tracking part of the algorithm.Considering the instability brought by the rapid matching of optical flow,the strategy of using multiple layers of image pyramid combination iterative optical flow is used to improve the stability of optical flow matching.(2)To address the problem of data error association caused by dynamic outliers of VSLAM algorithm in a dynamic environment,this paper proposes an outlier suppression mechanism combining reprojection error and iterative RANSAC based on the random sample consistency principle(RANSAC).The reprojection error is used to construct a separation model between inners and outliers,and the outliers are screened and eliminated before calculation to improve the robustness of the algorithm to the dynamic environment.(3)This paper designs a stereo visual SLAM algorithm,and implements the methods of improving real-time and enchancing robustness on this framework.The positioning accuracy of the algorithm is evaluated through public datasets to verify the effectiveness of the algorithm.(4)Based on the stereo visual SLAM,a visual-inertial odometry algorithm based on the theory of pre-integration integrating vision and inertial measurement by deriving IMU motion modeling and preintegration theory is proposed.Considering the instability of the fusion of vision and inertial measurement,an adaptive tracking strategy is designed,which selects pure visual tracking or visual inertial tracking according to the tracking state.By comparing the proposed pure vision algorithm and visual-inertial algorithm with public datasets,it is verified that the fusion of vision and inertial measurement effectively improves the robustness of pure visual SLAM for fast rotation.