Research On Visual Inertial Positioning Method In Dynamic Environment

Visual-Inertial Odometry(VIO)is a process of nonlinear optimization to estimate the pose state of intelligent robot in unknown environment.It uses the image feature information of binocular vision and the pose information of IMU.It is a prerequisite for intelligent robot to realize scene perception and autonomous navigation.The traditional VIO technology assumes that the scene is static.Once the object in the scene moves,the image will have imperfect information and motion blur,which will lead to the failure of inter frame feature point tracking,and then lead to the offset of sensor pose state estimation.In order to effectively solve the pose estimation problem of visual-inertial fusion in complex dynamic scenes,the following researches are carried out in this thesis:1.In order to obtain the spatial attitude information between the binocular vision and the IMU,a binocular visual-inertial space joint calibration algorithm is proposed.Firstly,the depth measurement principle and reprojection error are analyzed according to the projection model of the binocular camera,and the internal parameters of the binocular camera are solved by the chessboard calibration method.Secondly,the noise source in the adoption process is analyzed by the IMU measurement model,and the noise error is solved by the Allan variance.Finally,the transformation matrix of the camera and inertial navigation in space is used to solve the joint external parameter parameters,which improves the accuracy of using the IMU observation data to solve the camera rotation angle information.2.Due to the fact that moving objects in the actual scene will cause the occlusion of key information in the image.And lead to image feature matching errors.The accuracy of the vector pose estimation is reduced at last.A motion consistency detection algorithm based on L-K optical flow method and image semantic segmentation is proposed.While using L-K optical flow method to screen dynamic feature points,the convolution neural network based on Seg Net is used to complete image semantic segmentation;Match the dynamic feature points with the semantic segmentation graph to determine the moving object.Then eliminate the feature points in the moving object.Finally,the pose state is estimated by using the remaining static feature points.KITTI data set is selected for algorithm verification.The Simulation results show that the proposed algorithm has more accurate pose trajectory in dynamic scene than ORB-SLAM2.3.Because the visual positioning system is prone to the situation of no texture and fuzzy image,which leads to the failure of pose state estimation,a visual-inertial fusion positioning algorithm is proposed.Firstly,the IMU pre integration strategy is used to align the IMU information between image frames;Secondly,the objective optimization function is constructed by visual residuals and IMU residuals for nonlinear optimization.Finally,the loop detection module is used to complete the image similarity screening and pose relocation of the trajectory closed-loop area.It can effectively solve the problems of cumulative error and pose estimation offset caused by long-term operation of the system.4.An autonomous unmanned vehicle experiment platform is built.And the proposed in algorithm is verified in the actual campus scene.The experimental results show that compared with VINS-Fusion algorithm,the proposed algorithm is more accurate in pose estimation in complex large-scale dynamic scenes.