Research On Visual Navigation And Information Fusion Method Based On UAV

UAV system has a high demand for the stability and reliability of navigation systems.Due to the development of computer vision technology,the increase of computing speed and the reduction in cost have made visual navigation a reliable navigation method.The typical INS/GPS system requires satellite external information,which is susceptible to external influences such as electromagnetic interference,and the INS system has the disadvantage of errors accumulated over time.For the UAV’s flight characteristics and requirements for the navigation system,add an autonomous navigation method in the INS system,visual navigation,that does not rely on satellites.The combination of different navigation sensors can improve the navigation accuracy of the system,but it’s necessary to overcome the problems of different measurement frequencies of different navigation sensors,and the increase in the complexity of the system will also lead to a greatly increased probability of failure.It is necessary to improve the fault tolerance of multi-source information fusion methods.This article focuses on two aspects of UAV-based visual navigation algorithm and multi-source information fusion.The concrete research content is as follows:Image registration is the basis for visual navigation.First,several common feature extraction methods are derived in detail,including FAST corners,ORB features,and SIFT features.Among them,SIFT have advantages such as scale invariance and rotation invariance,but the stability is poor when the image affine occurs.In order to solve this problem,this paper gives the corresponding improvement methods.Considering that the MSA has better affine invariance,but the amount of calculation is large,so consider constructing the MSA operator based on the image block,and calculate the MSA feature vector in the circular area around the SIFT feature point.In order not to change the rotation invariance of the original feature,when calculating the MSA feature vector,the calculation range is proportional to the gradient amplitude of the center feature point.Combine the SIFT feature descriptor and the MSA feature vector to construct the SIFT+MSA combined feature,using the combined feature for image registration.In order to verify the image registration performance of the improved SIFT feature,use the image collected by UAV and the image artificially added with affine transformation for registration.The simulation experiment results show that compared with SIFT features,the SIFT+MSA combination feature and SIFT feature increase the computation cost,but the matching accuracy is improved.Based on the realization of accurate image matching,verify the feasibility of using only the visual sensor for motion estimation of UAV.The algorithm for obtaining UAV motion parameters based on sequence images using homography matrix decomposition is derived theoretically in detail.Using simulated UAV flight data to verify the algorithm,and analyze the influence of different Gaussian noises and different feature point distribution on the pose estimation results.The higher accuracy of feature point matching,the smaller the fluctuation of pose estimation error;the less relevant the feature points,the smaller the pose estimation error.Only using the vision sensor can estimate the position and attitude information of the UAV,but it will be limited by the movement status and external environment.Generally,the main navigation system is an inertial navigation system with strong autonomy,high update frequency and high estimation accuracy.Use visual navigation as an auxiliary navigation system to modify the inertial navigation results.The traditional loose-coupling algorithm independently estimate the motion of the two navigation systems of inertial navigation and visual navigation.The calculation is large and the fusion effect is general.This paper presents two tightly coupled algorithms based on vision/INS navigation-MSCKF and SWF.Fusion of the data in the visual sensor and the original data obtained by INS for estimation.Although this leads to increased calculation complexity,it can better extract the navigation information contained in the original data.Using the same sample set,compare and analyze the error characteristics of the pose calculation results of the two algorithms,then analyze the influence of the feature tracking length on MSCKF positioning accuracy.By comparing obtain the use scenarios of the two algorithms,SWF has better robustness,but the amount of calculation is much larger than MSCKF.MSCKF is more suitable for scenarios with limited computing resources such as UAV.In order to solve the problem of intermittent fusion of UAV multi-sensor integrated navigation system,and improve the adaptability of UAV navigation system to sensor failure,introduce the IMM theory into the federated filter.Verify the hybrid IMM’s ability to deal with sensor faults,comparing the processing capabilities of KF and residual?~2 fault diagnosis and hybrid interactive multi-model’s ability to deal with sensor faults,the simulation results show whether it is a sudden noise fault or Gradual failure,hybrid interactive multi-model algorithms have shown better performance.Based on the excellent fault tolerance performance of the hybrid interactive multi-model for noise,it is added to the federated filter,and the interactive multi-model is used in the sub-filter to replace the Kalman filter to improve the anti-jamming ability of the UAV integrated navigation system,The simulation results verify that the improved hybrid multi-model federated filtering can be effectively processed when a navigation sensor fails,reducing the impact of the less accurate sensor failure on the entire UAV navigation system.