Research On Image Matching Algorithm Of UAV Based On Improved AKAZE Feature

UAV remote sensing platform has become a research hotspot in the field of photogrammetry because of its flexibility,low cost and simple operation.And it has been deeply applied in environmental monitoring,map updating,disaster prediction,urban planning and so on.Due to poor stability of the drone platform,the imaging conditions are easily changed,which will result in a certain difference in translation,rotation,and viewing angle between the acquired images.These differences will make it difficult for later image matching processing.Therefore,improving the stability of the matching algorithm and obtaining the matching result with higher precision has become the key point of the matching processing of the UAV images.Based on the detailed analysis and discussion of the existing image matching algorithm,the UAV image matching algorithm is selected as the research object of this paper.Aimming at improving the algorithm adaptability,matching accuracy and time efficiency,studies on feature descriptor calculation,matching strategy and gross error elimination were made.The specific work carried out is as follows:1)Feature-based image matching algorithm.Analysis and comparison of feature-based image matching algorithms were made,including the SIFT,SURF,KAZE and AKAZE algorithm's theory and process.The comparison experiments in terms of time efficiency,accuracy and precision were made based on the above algorithms.2)Orthographic image matching of drone based on improved AKAZE feature.Based on the AKAZE algorithm,the inproved algorithm was proposed.First,the M-LDB descriptor was replaced by the OpponentFREAK descriptor in the feature descriptor calculation phase;Second,a staged matching strategy was used in the feature point matching phase to obtain high-precision matching results;Finally,in order to improve the time efficiency of the algorithm,the idea of parallel computing through OpenMP was introduced into two stages: feature point extraction and feature descriptor calculation.The experimental results show that the improved algorithm could achieve higher accuracy and higher precision matching results while improving time efficiency.Besides,its precision also could reach sub-pixel level.3)Oblique image matching based on Affine-AKAZE algorithm.In order to solve the problem that AKAZE algorithm had poor matching results for images with large viewing angle changes,the Affine-AKAZE algorithm which was applied to the registration of oblique images was psoposed based on the idea of ASIFT algorithm.The experimental results show that the algorithm had a good effect on image matching with large change of viewing angle.4)Image stitching and 3D point cloud extraction based on multi-view image.The algorithm of step 2 combined with multi-band fusion was used to complete feature point extraction and automatic stitching of UAV images.Besides,on the basis of Structure From Motion(SFM),the extraction of 3D point cloud data was realized.