Research On Optimization Technology Of UAV Images Positioning

As UAV remote sensing is widely used in national defense and economic development,the whole set of UAV image processing technologies are worth of deep study.High-quality and efficient UAV image positioning is the key step of the whole processing chain.This thesis researches on the optimization technologies for UAV image positioning,and focuses on the accuracy and robustness improvement strategies for feature extraction,image matching and block adjustment.Moreover,this thesis explores the parallel computation schemes of large-scale bundle adjustment.The effectiveness of the optimization technologies are verified by a variety of data sets.The main work and innovation points are listed as follows.(1)Firstly,an overview on the domestic and international research on UAV image processing is given,including the theories and technologies of feature extraction,image matching and UAV image's position and attitude recovery and optimization.And then the basic theory of parameter estimation problem in UAV image processing is deeply discussed and analyzed,including the commonly used algebraic methods,geometric methods,robust methods and statistical methods,as well as bundle block adjustment in photogrammetric and multi-view geometry in computer vision.(2)A hierarchical block strategy is designed for multi-view UAV images feature extraction and matching,which can take advantage of the GPU parallel acceleration,and help to alleviate the contradiction between large image frame and limited computational ability.At the same time,the Union Find Set in graph theory is utilized to realize the rapid extraction of corresponding points,and further to enhance the intensity and stability of bundle adjustment network.The efficiency of proposed strategies and methods are verified by multiple sets of UAV image matching test.(3)For desert area images with low contrast,repetitive patterns and uneven texture,problems of scarce tie points or large number of mismatches are significantly serious.To solve these problems,a texture adaptive method for tie point extraction is proposed for desert area.Firstly,the image pyramid is built to perform initial matching and establish the approximate homography transformation between the images.Then,the texture level is evaluated based on the grey level covariance matrix of the image blocks,and the adaptive extraction and matching of the feature points is undertaken.Experiments show that the proposed method could obtain enough,reliable and homogenous corresponding points.(4)To solve the problem of mismatches and large water coverage in the island and reef images,a method based on the false alarm minimization is proposed to eliminate mismatches.The number of false alarm from random samples is counted up according to the residual distribution,and then the thresholds for inlier and outlier is determined adaptively in order to calculate the accurate geometric constraint model.Experiments on various typical UAV island and reef images demonstrate that the method is necessary and effective.(5)In order to improve the positioning accuracy of UAV images,the POS placement error,camera interior and exterior orientation parameters and lens distortion are calibrated and compensated.Structure From Motion is introduced into UAV image positioning,and comparative study between SFM and POS-aided bundle adjustment is undertaken.Results show that these two methods can achieve comparable accuracy.Therefore,POS-aided bundle adjustment is preferred for the direct geo-referencing in emergency conditions,whereas SFM can be used in normal tasks to ensure accuracy and reduce costs.(6)In normal UAV images block adjustment,the standard L2 norm is sensitive to noise,especially the Levenberg-Marquardt method,which is used to overcome the ill-conditioned state of normal equation.To overcome these difficulties,a robust adjustment method in consideration of the observation's reliability is proposed,which can tolerate high level noise.The method can adaptively adjust the cost function according to the overlap and the residual of the feature points,thereby overcome the impact of the mismatches on the adjustment results.In order to evaluate the effect of robust adjustment,a corresponding precision evaluation method is put forward.Experiments show that the proposed method in this thesis is robust and stable even in dealing with a high level of noise.(7)In order to enhance the computational efficiency of large-scale bundle adjustment of UAV images,a parallel optimization method based on preconditioned conjugate gradient and inaccurate Newton solution is designed.This method can avoid the direct inversion calculation of large-scale equations and reduce the number of iterations for conjugate gradient.To accommodate CPU multi-core and GPU parallel computation,the sparse matrix is stored in blocked compressed sparse row format to save the RAM resource requirements on CPU and GPU.Experiments on multi-batch UAV images show that the proposed bundle adjustment scheme could process more than 10,000 images on a regular-configuration computing platform with high efficiency and speed-up ratio.