Research And Application On Matching Algorithm Of Unmanned Aerial Vehicle Image

Unmanned Aerial Vehicle(UAV)has the characteristics of solidity,which can quickly and safely obtain ground images even in complex environment,and is widely used in agriculture,industry and military fields.Because the information of a single UAV image is limited,it cannot reflect the whole target area,and it needs to be stitched into a panoramic image.Among them,image matching occupies most of the time of image stitching,which makes the image stitching process long and inefficient.To solve these problems,based on the scale-space theory,this thesis is studied from three aspects: scalespace construction,feature point detection,and feature point description.The main work is as follows.(1)To solve the problems of large approximate discrete error of Gaussian convolution kernel and low utilization rate of scale layers in the traditional Difference of Gaussians(DOG)approximate Laplacian of Gaussian(LOG)method,a method of subtracting several adjacent Gaussian scale layers is proposed to construct the DOG scalespace,and the inconsistent amplitude of the DOG response value is normalized.The correctness of the scale normalization method is proved by experiments.Compared with Scale-Invariant Feature Transform(SIFT),Speeded Up Robust Features(SURF),Oriented FAST and Rotated BRIEF(ORB),Binary Robust Invariant Scalable Keypoints(BRISK)and other detectors in Oxford data sets and UAV images,the test results show that the scale-space constructed by the suggested algorithm has higher repeatability.(2)Aiming at the ill-conditioned problem of Taylor iterative accurate location of extreme points,an accurate location algorithm of extreme points based on least squares is proposed.Firstly,the ternary quadratic equation is fitted according to the sample points in the neighborhood of the pseudo-extreme point 3×3×3.Then,the least square method is used to solve the coefficients of the equation.Finally,the accurate position of the extreme point is obtained by solving the first partial derivative of the equation,and unstable points are eliminated according to the discriminant conditions to improve repeatability.In order to speed up the operation,the neighborhood coordinates of all pseudo-extreme points are set to the same coordinates,thus reducing the fitting process of the equation.The experimental results show that the average time consumption of the method before optimization is 3.18 times that of Taylor iteration,and that of the optimized method is 0.68 times,and its repetition rate in various scenes is obviously improved.(3)Aiming at the problems of low robustness,large storage space and sensitivity to noise,a binary descriptor based on Sobel gradient comparison is proposed.Firstly,concentric circles with radii of6σ,11σ and15σ are set at the center of the feature point,and each circle is divided into 8 regions in angular direction.Then,in order to reduce the influence of noise,the Sobel operator is used to calculate the gradient value of each area,which is divided into eight gradient directions.Finally,a binary descriptor is generated by comparing gradient values.Experimental results show that the developed algorithm has better robustness,smaller storage space and faster matching speed than SIFT,SURF,KAZE,Fast Retina Keypoint(FREAK)and other descriptors in UAV images.