Research On Localization Algorithm Based On Vision-Inertial Odometry

Localization technology is one of the key technologies of mobile robots,and precise localization system is the basis of control and planning.Both vision and inertial navigation can get relative localization result.Vision have rich external environment information,but the accuracy of localization is easy affected by the environment and image quality.Inertia localization is easy to be affected by the cumulative error caused by the integral.Therefore,use the advantages and complementarity of two sensors to improve the accuracy and robustness of the pose estimation is the main content of this dissertation,which is divided into the following four aspects:(1)Visual inertia fusion algorithm separate the attitude and position estimate.In the existing IMU and vision fusion algorithm,the attitude and position are usually regarded as the state variable together,which will lead to the divergence when solving the model.In this dissertation,the position and attitude are estimated separately.The attitude is estimated by IMU,which is more precision and robustness at the same time avoids the initialization of the camera attitude at the begin.(2)Using the error state Kalman filter model estimate the IMU attitude.The error state transforms the filter from a strongly non-linear attitude space to a near-linear attitude space,avoiding errors introduced by the linearization process.Meanwhile because of the attitude solution in the vicinity of the error state,the error angle parameter value of the state estimation is small,which avoids the parameter singularity as the gimbal lock problem.By using the data of magnetic compass and accelerometer as observation,the attitude of gyro integration is corrected by gravity reference vector and geomagnetic reference vector,and the attitude estimation accuracy of the system is improved.(3)Based on the multi-state constraint Kalman filtering method,the vision information is used to estimate the position error.The feature processing is delayed until they are out of view,which avoids the usual problems associated to feature initialization,and the feature point position is estimated by using the Gaussian Newton optimization with inverse depth parameter.State estimation utilizes the geometric constraints of the features observed without adding them to the state,so the algorithm has linear relationship with the number of features,which makes it suitable for real-time execution.(4)Close loop detection and global optimization is used to solve the problem of localization drift.The estimated localization result with IMU and vision will have some drift.The global optimization using the boundle adjustment method with close loop frame detected by the bag of words model can effectively solve the drift problem.