On The Key Techniques For Visual-Inertial Tight Coupled Navigation Systems Of UAV

As an intelligent aerial robot capable of autonomous flight and remote control,UAV is widely used in various military and civilian fields.The accuracy and reliability of its navigation system are critical to the success of UAV missions.The Visual/Inertial Tight Coupled Navigation System(VITCNS)has become an important research trend of UAV navigation systems due to its superior sensor complementarity and high-precision.This dissertation focuses on the accuracy and reliability of the VITCNS,and the modeling of VITCNS,the offline calibration of camera and Inertial Measurement Unit(IMU),the adaptive online calibration of IMU,the initialization and mismatched features elimination technology are studied.The main research work and results are as follows:1.The model of VITCNS is designed.The measurement and error model of camera and IMU are established,and the LK optical flow tracking method based on the image pyramid are studied to track the features,and the judgment condition for the quality of features is analyzed.To avoid the out-sync of camera and IMU,the IMU pre-integration algorithm is derived with the weighting method to process the pre-integration values of the IMU measurements adjoining the image,realizing the alignment of the camera and IMU measurements.The solution of camera poses based on matching points between adjacent frames is studied,including the epipolar geometry,Pn P and ICP algorithms,to build the model of VITCNS.2.The offline calibration technology of camera and IMU are studied.To avoid the step-like ouputs of IMU and not applicable with the complex temperature field based on the traditional IMU calibration,the temperature compensation based on the multiple regression and BP neural network are proposed.The IMU nonlinear scaling factors are calibrated to improve the dynamic performance of IMU,and the calibration method based on Kalman filter(KF)and Multi-position Nonlinear Optimization(MNO)are proposed,with high accuracy and low requirements for equipment.Besides,the calibration for IMU random errors based on the Allan variance model are studied,and the intrinsic parameters and distortion coefficients of camera,the transformation matrix between camera and IMU are calibrated using Kalibr toolbox.The semi-physical experiments prove that the performance based on BP neural network is superior.The MNO estimates all parameters without relying on the equipment’s accuracy and adjusting the parameters cumbersomely.3.The adaptive online calibration technology of IMU for VINS are studied.Due to the complex and changeable working environment of UAVs,the IMU parameters are susceptible to temperature,humidity,vibration,etc.,hence an online calibration method for IMU biases,scale factors and misalignment errors is proposed.The continuous and discrete IMU pre-integration models with IMU parameters are derived,and the residuals are added to the cost function,hence the online calibration becomes an optimization problem.By deriving the Jacobian matrix of the residuals,the cost function can be solved by nonlinear optimization.The regularized mahalanobis distance is proposed to solve the singular problem of the covariance matrix,and the adjust parameters that reflect the intensity of the motion are introduced to adjust the optimization weights adaptively for IMU parameters.In addition,the adaptive online calibration method could run in real-time and is verified using the public datasets.The results prove that the proposed method greatly improves the navigation accuracy compared with the non-adaptive and VINS-Mono method.4.The initialization and image matching algrithom of the monocular VITCNS.The initialization based on loose coupled method is studied,and the camera measurements are processed by the SFM algorithm to initialize the poses and landmarks.Besides,the gyro biases,gravity vector,scale factor initial velocity are estimated according to the results of SFM and pre-integration.The LK optical flow tracking method with elimination of mismatched features based on the IMU pre-integration is proposed,to eliminate the mismatched features between adjacent frames caused by the large motion of UAV,which makes the IMU pre-integration results as the motion constraint to predict the features’ positions.By comparing the preditions with the tracking positions,the features whose errors exceed the threshold are eliminated.Besides,the optimization model for visual residuals are designed,and its Jacobian matrix are derived to complete the overall cost function of VITCNS.Experiments on the public datasets verify the effectiveness of elimination method,thereby improving the accuracy of navigation.5.The design and experiments of VITCNS.The gazebo-based UAV simulation system,Singleaxis Rotation Inertial Navigation System(SRINS),and platform based on Android mobile phone are designed to verify the initialization,offline and online calibration,elimination mismatched features and optimization estimation algorithms for VITCNS.The experiments of initialization is performed with different motions in the UAV simulation system,and a motion for initialization with fast speed and high-precision is designed.In addition,the results prove the accuracy,speed and navigation results of the adaptive online calibration is better than that of non-adaptive method.The IMU temperature compensation and the MNO are verified in the SRINS,and the quasi-static,swing state and vehicle experiments prove the navigation precision is improved.The experiments of elimination and online calibration method are performed on the phone platform,and the two methods achieve the highest navigation accuracy.