Research On Visual-inertial Odometry Based On Initial Stage Optimization

Till now,although the research on localization and navigation technology has made a lot of achievements,for high precision positioning technology,these methods are still insufficient.Due to the limitations of a single sensor,it cannot locate accurately in some scenes.It is necessary to fuse measurements from multiple sensors to obtain more robust initialization information.The typical one is the visual-inertial odometry system,which fuse a monocular camera and an inertial measurement unit(IMU).The initialization accuracy of visual-inertial odometry system directly affects the performance of entire SLAM system.As the data fusion process of the two sensors will generate errors due to their characteristics,and the performance of visual-inertial odometry system depends on the vision module initialization.This paper studies the fusion algorithm of monocular camera and IMU in initialization stage of visual-inertial odometry.This thesis mainly studies in the following aspects:(1)Visual-inertial odometry system with extrinsic parameters optimization simultaneously.This method can estimate the extrinsic parameters of the monocular camera and IMU online,and obtain other measurements at the same time.The rotation extrinsic parameters are estimated by matching the pose information obtained by the vision and the IMU module.After the gyroscope bias is estimated,translation extrinsic parameters are added to the optimization variables,the scale factor is estimated and gravity vector is refined.Finally,the velocity is calculated through the estimated measurements and the accelerometer bias is processed in the back-end nonlinear optimization.Experimental results on public datasets verify the effectiveness of this method.(2)Visual-inertial odometry system with online temporal offset optimization.This method can estimate and compensate the temporal offset between monocular camera and IMU online.Assuming that the timestamps of the IMU are correct,and one feature point in space is observed by two consecutive camera keyframes at the same time.Due to the effect of temporal offset,the coordinates of the feature point have changed.By optimizing the reprojection error of the changed feature points on another camera keyframe under IMU constraints,the temporal offset and other measurements are estimated.Finally,the timestamps of camera are aligned with that of IMU according to the estimated temporal offset.The experimental results of public datasets and realworld scenes demonstrate that this method can improve the accuracy of localization.(3)Visual-inertial odometry system with point and line features simultaneously.This method extracts point features and line features combined with IMU measurements for pose estimation.The time consuming and computational complexity of line features extracting are relatively high.By eliminating short-line features,the time for extracting the line features is reduced to ensure real-time performance.As line features can make better use of planar characteristics,homography and fundamental matrices are used for pose estimation in planar and non-planar scenes,respectively.Finally,the pose information obtained by the vision module is fused with the IMU measurements to obtain more accurate initialization information.The experimental results of public datasets prove that the proposed method can improve the robustness of camera initialization and the pose estimation accuracy of the entire system.