Research On Fast And Reliable Image Matching Algorithm For Staring Satellite Image Stabilization

Staring satellite is a new type of earth observation satellite which has been developing in recent years.Through staring and imaging in a certain area,it gets timing sequence images,which are suitable for analyzing moving targets and obtaining instantaneous characteristics of targets.These important dynamic information is difficult to get from traditional earth observation satellites.There is geometric offset between sequence images,and moving targets between images affect the accuracy of relative geometric relations.Also,there are some errors in the absolute positioning of the images.These are not conducive to the detection,extraction and recognition of the moving targets.Meanwhile,fast registration can promote the rapid application of staring satellite.Image matching is a good way to eliminate the relative and absolute position offset of staring satellite images.In this paper,we study how to fast and reliably eliminate relative and absolute position offset of staring satellite images by the method of image matching,which is one of the key problems of remote sensing to earth observation.In view of that relative and absolute position offset of staring satellite sequence images,the paper studies the image stabilization algorithm base on image matching,including the sequence image matching algorithm,the error matching point elimination algorithm for moving target,the motion compensation algorithm for relative and absolute position offset,and the development of embedded GPU stability imaging system.These provide a new solution to the rapid application of staring satellite,to the masses and to serve public,which has important guiding significance and application value.The specific research contents and main innovations are summarized as follows:(1)A sequence image matching algorithm based on multi-thread adjacent region ORB feature is proposed.SR-SIFT is performed on the down-sampling images to obtain the rough geometry relation between images.According to the rough geometry relation between images,the ORB features only search areas where the match point might exist,which can reduces mismatches,and improve the reliability of the ORB feature match.First,the large size satellite image is divided into blocks,and the corresponding block is establish.Then,the adjacent region ORB is performed in each block,the accuracy of matching results are improved with least square matching.Finally,the algorithm is parallelized with OpenMP.(2)An algorithm of eliminating mismatching points on the moving targets is proposed.Firstly,according to the consistency of motion direction,part of matching points on the moving target are eliminated,also the correct matching rate is improved.Then,the non-parametric statistical rank is used to detect the cloud edge,the mismatching points located at the edge of cloud are remove.Finally,a coarse to fine RANSAC algorithm is adopted to eliminate mismatching points.In coarse RANSAC,the rigid transformation is used,then in the fine RANSAC,the rigid transformation is used.(3)A motion compensation algorithm based on the control point matching is proposed to compensate the relative and absolute position offset of staring satellite images.The control point matching is perform on reference frame,and the absolute geometry relationship of reference frame is obtained.Then,the motion compensation parameters of each frame are established according to the motion geometry relation of the frame and the absolute geometry relationship of reference frame.For robust control point matching,the coarse registration,the pyramid relaxation and the position offset clustering mismatching eliminated method are adopted.(4)The embedded GPU image stabilization system is developedThe parallelism and load in the algorithm steps are analyzed.And according to the characteristics of the embedded GPU,the algorithm of staring satellite image stabilization is migrated and optimized.On the above research contents,sequence images of GF-4 and Jilin-1 are selected to verify the reliability of the proposed algorithms.The experimental results show that the sequence image matching method algorithm based on multi-thread adjacent region ORB feature has low memory and fast speed,and has good robustness against radiation difference and angle difference.this article selects GF-4 and Jilin-1 gaze sequence of experiments have been carried out to verify the above image matching algorithm,the experimental results show that in this paper,the sequence of image matching algorithm has a low memory,speed faster,radiation differences,the differences of perspective has better robustness.The mismatching points on the moving targets can be eliminated by the proposed mismatch points eliminating algorithm.And the stabilization accuracy is improved.The control point matching motion compensation can eliminate the relative and absolute offset of staring satellite images.Finally,the stability image algorithm of this paper is developed with embedded GPU,the efficiency of image stabilization is improve.The proposed stable method based image matching has solved the relative and absolute offset of staring satellite images.After stable image processing,the staring satellite images have good relative position consistency and good geographical accuracy.It is beneficial to the target detection,and can be superimposed directly with the information of road,place name and other attribute information.It is of practical significance to have a good intelligent analysis of the observation target and promote the application of the time sequence image of staring satellite.