Research On Pose Estimation Method Bsed On IMU And Monocular Vision Fusion

As one of the most promising research directions in the field of visual SLAM,visual inertial odometry(VIO)has attracted more and more scholars' attention.Low-cost,miniaturized Inertial Measurement Unit(IMU)and camera are combined by VIO.On the one hand,it constrains the drift of IMU by using the rich information of the environment captured by camera,on the other hand,it recovers the metric scale of monocular vision and be aligned with inertial frame by using the self-motion information of the carrier measured by IMU,the complementary nature of the two sensing modalities is utilized for navigation and localization.Unlike with the traditional filtering method,this paper adopts a tightly coupled scheme based on non-linear optimization with higher localization accuracy to fuse sensor information.The main research work of this paper is summarized as follows:1.The coordinate system involved in VIO system is defined and the frame transformation relationship is clarified.The IMU measurement model and pinhole camera projection model are analyzed,and the mathematical model of visual SLAM problem is established,and the solution method based on non-linear least squares is introduced.2.The initialization method of VIO system is researched.The image features are tracked by KLT optical flow method at the front end of the vision system,and the camera pose is recovered by SFM.At the same time,the IMU measurement between two consecutive frames are pre-integrated.Finally,the poses obtained by IMU and vision are loosely coupled.Furthermore,the initial values of the system for back-end optimization,including the metric scale,gyroscope bias,direction of gravity acceleration and velocity,are obtained by solving the linear least square equation3.The cost function of non-linear optimization is constructed by IMU measurement error and visual reprojection error.At the same time,the strategy of "key frame" marginalization is adopted to ensure the constant computational consumption and real-time performance of system.In order to avoid the loss of information,the prior information from marginalized is added to the cost function.In addition,loop detection and Four-Degree-of-Freedom pose-graph optimization are added to reduce the cumulative error of the system and ensure the global consistency of the system trajectory.At the end,the paper presents an attitude estimation method based on Extended Kalman Filter under the motion condition of VIO pure rotation4.The VIO localization algorithm is implemented on ROS robot operating system,and the actual hardware test and dataset test verify the robustness and effectiveness of the approach.The localization effect of the algorithm is analyzed qualitatively and quantitatively.The results of experiment show that the loop closure significantly improves the localization accuracy of the visual-inertial navigation system and reduces the accumulated drift.For pure rotating motion scenes,the attitude estimation algorithm based on Extended Kalman filter can also achieve a high attitude estimation accuracy.