Study On Inertial/Vision Integrated Navigation System

Since strapdown inertial navigation system?SINS?has accumulated errors with the time progress,other sensors like GPS are expected to aid SINS in navigation.It is known that GPS dropouts easily when the signal is shielded or interfered.This paper is devoted to study the vision aided SINS integrated navigation system,to solve the problem that how to navigate through the camera in visual navigation system?VNS?and how to do the data fusion in the integrated navigation system.Visual odometry is a kind of simple and convenient sensor that navigates by processing the images from the camera.According to the number of the camera,it contains monocular,binocular and multi-eye visual odometry.Compared with other visual odometry,monocular visual odometry has low cost and simple design,so it is applied in this work.Extracting features from images is an essential part in visual navigation.David Lowe proposed scale-invariant feature transform?SIFT?which can distinctive image features from scale-invariant key-points.However,SIFT is computation-intensive and time-consuming.This paper set a threshold to limit the numbers of key-points so that the running time becomes shorter.In order to improve correct match rate,both distance measurement and comparability function are applied in similarity measurement.The matching point is used to calculate the fundamental matrix by eight point method.After doing the singular value decomposition of the essential matrix,the relative position and attitude are obtained.In the SINS/VNS integration system,the estimate errors of SINS are chosen to be the state variables,and state equation is established based on the error equation of SINS.The observation is the difference of SINS's relative position and attitude and VNS's relative position and attitude.Kalman filter is used to do the data fusion.To achieve the optimal estimate,Kalman filter is under the assumption that state noise and measurement noise are supposed to be Gaussian white noise.Also the model parameters should matched with the actual condition.However,there are some uncertainties when the integrated navigation system is in application,so single model cannot represent the actual system since the model parameters or noise are time varying which will obtain an unsatisfied estimations.As for the uncertainties,multiple model estimation can achieve a optimal estimate by selecting a set of sub-model to describe the variable system.And the interacting multiple model?IMM?is one of the multiple model estimation which have a better performance and can adaptively adjust by the actual environment.In addition,the Huber method is a kind of estimator which combinesl₁/l₂norms.This approach is robust to the uncertain noises especially the non-Gaussian noises caused by outliers.Then,the sub-filter of IMM method is set to be Huber-based Kalman filter,and the proposed IMM-Huber method is robust to the uncertain and contaminated noises.Therefore,the application of IMM-Huber method can settle the problem of uncertainty and improve accuracy of the integrated navigation system.