Study On Binocular Visual Inertial Odometry Of Multi-pose Information Fusion

With the development of machine vision,inertial navigation and computer science technology,the visual-inertial odometry is used to solve the problem of continuous,stable and accurate positioning and attitude measurement of the vehicle under the condition that the GNSS signal strongly rejecting the weak observation in the unstructured scene.The visual-inertial odometry has gradually become a hot research direction in the field of navigation.However,the performance of a visual-inertial odometry is related to scene characteristics,vehicle motion,and performance of sensors.The following situations may occur during visual-inertial odometry work:Geometric transformation between images caused by camera viewpoint changes;insufficient feature information and illumination changes in the scene;motion blur caused by high dynamics of the vehicle;noise of the camera detector.All above factors will cause the pose measurement of the visual-inertial odometry to be larger cumulative error and even invalid.Due to factors such as zero-bias instability and noise of the inertial measurement unit,the position,velocity and attitude of inertial navigation will also produce cumulative errors.Therefore,the accuracy and applicability of the visualinertial odometry relative to other navigation systems is greatly affected by scene characteristics,carrier motion,and sensor performance.This paper considers that the binocular visual odometry has two monocular and one stereo vision pose measurement information,which provided more observation dimensions for the visual-inertial odometry.The integration of the information with the inertial navigation will improve the accuracy of pose estimation of visual-inertial odometry.Moreover,the multi-visual pose measurement information can provide more measurement margins.When a visual pose measurement unit diverges or fails,other visual pose measurement units and inertial navigation information can still obtain stable navigation information with higher precision.The purpose of this paper is to study a fusion algorithm and redundancy structure of multi-pose measurement information of binocular visual inertial odomtery,which makes the visual-inertial odometry higher robustness,accuracy and applicability in the condition of geometric transformation,feature sparseness,illumination variation,blur and noise,etc.Depending on algorithm and structure,stable,accurate and real-time autonomous navigation result will be realized for visual-inertial odometry.The main research contents of this paper include:1.This paper proposed a visual-inertial odometry with the multi-visual pose information fusion working mode.This paper systematically analyzed the complementarity of the monocular visual odometry and the stereo visual odometry pose estimation algorithms,and proposes the "double monocular + stereo" binocular visualinertial odometry working mode.The binocular visual inertial odometry has the advantages that the pose estimation is not limited by the depth of landmarks and an unambiguous scale factor.In the back end,the landmark information with more quantity,higher triangulation measurement precision and better geometric distribution is utilized to improve the estimated precision.With the datasets testing,the "double monocular + stereo" working mode has improved the translation accuracy of the visual inertial odometry by 27.34% and the rotation accuracy is improved by 13.54% compared with the stereo working mode.2.In this paper,the redundant structure of visual-inertial odometry with multi-pose information is proposed.In order to solve the situation of large-scale conditions that sparse features,feature mismatch,VO pose estimation algorithm divergence,based on the "double monocular + stereo" working mode,the motion state calculated by IMU pre-integration is used for the fault monitoring VO pose results.Once a visual odometry unit fails,the fault unit can be automatically initialized according to the state machine,the navigation process can still be realized by the normal VO unit and INS.Under the large-scale scene conditions,the visual-inertial odometry function structure with redundant design of visual pose information significantly improves the robustness of the visual-inertial odometry.3.This paper proposed an image feature matching algorithm aided by IMU preintegration.In this paper,the relative pose between cameras predicted by IMU preintegration is used to estimate the region of the corresponding feature location in the matching image according to the epipolar geometric constraint between images,so that the search region of the corresponding feature is limited to the neighborhood of the corresponding epipolar line in the matched image.range.Combined with the IMUassisted 2-point RANSAC algorithm,compared with the classical image feature matching and 3-point RANSAC algorithm processing time,the image matching processing time of this paper is reduced by 34.02%,which made up for the "double monocular + stereo" working mode.4.This paper proposed an improved IMU pre-integration objective function model.The back-end of the binocular visual inertial odometry used a batch smoothing optimization method to fuse the 6DOF motion information of the three visual-odometry units and inertial navigation.In order to ensure that the motion increment of the inertial navigation calculation is consistent with the time of the VO pose increment,the objective function of the inertial navigation is established in the form of IMU preintegration.Different from the classical IMU pre-integration model in the navigation system,this paper established the pre-integration model of IMU based on the body coordinate frame,and derived its first-order form based on the assumption that the IMU zero-bias was slow-variable drift,and established the INS based on the body frame.The objective function derived the Jacobian matrix of the objective function relative to the state variable.The improved IMU pre-integration model based on the body frame avoided the iterative iteration of the rotation matrix of the body frame and the navigation frame in the back-end estimation process,and the Jacobian matrix also has a more compact form.With the improved IMU pre-integration model,the amount of calculation of the back-end optimal estimate was reduced.In this paper,the performance of the binocular visual-inertial odometry in indoor scenes and large-scale scenes is verified by dataset and vehicle test.The indoor scene utilizes the EuRoC dataset.Compared with the OKVIS visual-inertial odometry,the translation accuracy of the binocular visual-inertial odometry is increased by 18.37%,and the rotation accuracy is improved by 10.52%.In the large-scale scene,the highprecision GNSS/INS integrated navigation system output is used as the ground truth,and the binocular visual-inertial odometry is tested and verified.The test results show that the binocular visual odometry has high robustness,invariance and real-time.Under the condition of 7.2km motion distance,the translation accuracy is 1.35%,the rotation accuracy is 0.0020 /m and the data rate is 10 Hz.