Research On Image Matching Algorithms For Vision-aided UAV Localization

Vision-aided UAV localization refers to the use of combination between the UAV aerial image information and the geographic information system in prior knowledge,such as the satellite remote sensing image between the geographical indication and the localization,and the method to estimate for the pose of UAV itself.The vision-aided UAV localization not only is becoming the powerful supplement to global navigation satellite system(GNSS)and inertial navigation system(INS)technology,but also is playing important supporting role for improving of the interference ability in UAV navigation system.The image matching algorithm is one of the key issues in the field of the vision-aided localization technology.With some conditions to exist between the UAV aerial image and the satellite remote sensing image,such as lighting condition,visual angle,dimension and the difference among distributed ground object,the design of the image matching algorithm faces very challenges with speediness,robustness and accuracy.The research contents include feature extraction,target detection,searching,recognition,camera calibration and so on,and the research,which has the important theoretical research and application promotion value,can be widely applied to UAV visual navigation,identification and tracking of ground object,satellite image recognition,and land use land cover detection.This dissertation makes research background about the development of the vision-aided UAV localization system,makes the key technology of the UAV localization based on feature matching as major research direction,and also,makes the improvement of robustness and anti-interference ability in the UAV localization technology as the research purpose.With the design of local and global feature descriptor,the feature matching strategy,the global feature-based quick search of target images,the fusion method between INS system and vision-aided localization,the dissertation proposes the image matching algorithm-based vision-aided UAV localization strategy for a set of relatively complete,practical and more superior performance.The main research contents and results are reflected in the following respects:First of all,to improve invariance,real-time and significance of the gradient distribution-based local feature descriptor,a new gradient distribution statistical model is proposed.With the gradient magnitude accumulation orientation histogram as the discrete marginal integration function,the nonparametric estimator is used to realize the accurate estimate of the function.On the very important rotation of the image matching algorithm in the UAV visual navigation and localization application,and the problem of contrast and scale invariant,the feature descriptor,the reduction of the feature vector dimension,and the computational complexity based on the radial sampling grid are proposed.The feature description can be applied to the local feature descriptor based on one order gradient histogram.The algorithm which has small amount of calculation is simple and the real-time of application at the graphical processing unit or the imbedded devices can be ensured.Compared with other algorithms,the experiment results show that the proposed feature descriptor has better robustness and effectiveness.Second,on the problem of local feature matching accuracy in the process for vision-aided UAV localization,the feature matching algorithm based on both the adaptive near neighbor distance rate and the gradient predominate orientation difference is proposed.In the first,to improve matching efficiency,with increasing the number of matching,the adaptive near neighbor distance rate of the number of feature point based on the satellite image is proposed,and then the corresponding position of UAV image in the image data base is estimated analytically by using the gradient direction of the main deviation statistics.Finally,by using GMM model and EM algorithm,the distributed probability density function is estimated at the gradient direction of the main deviation statistics and then the main direction of deviation angle already acquired is applied to the matching process.By comparing with other main matching algorithm,the three kinds of data base image,which is selected from the experimental stage,are presented to validate the effectiveness of the proposed algorithm.Also,on the problem of the UAV localization at complex urban environment,a global feature descriptor of the remote sensing image and the matching algorithm based on line segment extraction is proposed.With considering of distributed artificial layout features such as building,road,and vegetation region in aerial image about urban region,a Line segment extraction algorithm based on PCA is proposed for the feature to overcome the change of the lighting and contrast.The feature based on line segment is satisfied the rotation invariant,so that it can be used for robustness of matching between aerial image and satellite image.Finally,by using hierarchical search method based on the global feature,the UAV aerial image is identified from the reference image database constructed with the small region satellite images.The experiment results show that the search and matching for the UAV aerial image could be realized quickly and effectively by using the proposed method and the robustness of the localization algorithm could be improved under the variation of both rotation and light condition.Finally,on the problem of inertial navigation and vision-aided localization,the UAV localization algorithm flow chart based on the fusion between global feature and local feature is proposed.The algorithm flow uses global feature matching to process the rough estimation about the UAV aerial image localization.Then,by using GMM estimating algorithm on the gradient direction of the main deviation,small regional satellite images which is similar to the UAV aerial image is obtained.Finally,the accurate estimation of the UAV coordinates is realized by using ORSA algorithm.Experiment results of global feature matching are presented to validate super performance of both region search and image matching in proposed algorithm.In conclusion,a complete vision-aided UAV recombination localization strategy based on both INS and image matching is proposed.On the defects of existing methods,a series of the matched algorithm between local and global in the feature descriptor and point and region searching algorithm are proposed to improve the accuracy,robustness and real-time for vision-aided UAV localization.Research results have important theoretical significance and application value.