Research On Aerial Image Matching Algorithm Of Complex Terrain Based On Point Feature

With the rapid development of surveying and mapping technology,UAV aerial photogrammetry has become one of the indispensable methods for obtaining spatial data of surveying and mapping geographic information.Image matching is a key prerequisite for the production and application of related image data such as image stitching and 3D modeling.In recent years,the related algorithms of image matching technology have been quite mature with the joint efforts of related researchers.However,for complicated terrain,the complex terrain with many undulations,occlusions,shadows and small grayscale changes brings some difficulties to image matching.And these making some previous algorithms have fewer correct matches,poor real-time performance,and fewer feature points.In view of this,this article first analyzes the characteristics of complex terrain imagery and its impact on matching;secondly,based on the two angles of nonlinear diffusion filtering and FAST corner points,an improved AKAZE algorithm based on DAISY descriptor and an improved BRISK algorithm based on LATCH descriptor which is based on machine learning.After comparing and analyzing the two algorithms with the previous algorithms,the improved AKAZE algorithm and the improved BRISK algorithm are compared and analyzed.The main research contents and results of this paper are as follows:(1)On the basis of learning the related theories of image collection,the paper analyzes and summarizes the characteristics of UAV aerial images in complex terrain areas and the difficulties in image matching.The characteristics of UAV aerial images in complex terrain areas are mainly multi-terrain undulations,mountain occlusions,small pixel grayscale discrimination,and large image geometric distortions.These characteristics make it difficult to identify features,poor feature robustness,and easy mismatches during the image matching process.(2)Explained and analyzed the algorithm based on linear Gaussian filtering,the algorithm based on nonlinear diffusion filtering and the algorithm based on FAST corner points respectively.At the same time,the three types of algorithms are compared and analyzed.The algorithm based on linear Gaussian filtering has better robustness,but has fewer feature points and general timeliness.Therefore,this type of algorithm has slightly insufficient adaptability in complex terrain;The adaptability of the algorithm based on nonlinear diffusion filtering in complex terrain areas has been improved,but the KAZE algorithm has poor timeliness,while the AKAZE algorithm has improved the timeliness,but the feature robustness has also decreased;The algorithm based on FAST corner points has its advantage in timeliness,but its robustness is lacking,and its adaptability to complex terrain images is general.(3)In view of the lack of robustness and long time-consuming features of the AKAZE algorithm in the UAV image matching of complex terrain,This paper constructs an improved AKAZE algorithm by introducing the DASIY descriptor.The algorithm first uses the AKAZE detector to perform feature detection,and then uses the DAISY descriptor to describe the feature,then uses the FLANN algorithm to perform rough feature matching,and finally uses the random sampling consensus algorithm based on the homography matrix to eliminate mismatches to complete the fine matching.Experiments on three sets of complex terrain images show that the improved AKAZE algorithm in this paper maintains the feature detection ability equivalent to the AKAZE algorithm,and has higher real-time performance and more correct matches than the SURF algorithm and the AKAZE algorithm.(4)Aiming at the low utilization rate of feature points and low real-time performance in the BRISK algorithm in the UAV image matching of complex terrain,this paper proposes an improved BRISK based on the descriptors of neighborhood gradient convergence and traditional binary descriptors.algorithm.The algorithm uses BRISK detectors to detect features,uses machine learning-based LATCH descriptors for feature description,FLANN algorithm for rough matching,and finally a random sampling consensus algorithm based on homograph matrix to complete fine matching.Experiments on three sets of complex terrain images show that the improved BRISK algorithm in this paper obtains more correct matches than the BRISK algorithm and the SURF algorithm when the total number of matches is slightly lower than that of the BRISK algorithm,and the single correctness is time-consuming.Optimal,the algorithm is more efficient.(5)A comparative analysis of the improved AKAZE algorithm and the improved BRISK algorithm is carried out.The experimental matching result graph and related index statistics show that the two algorithms have their own advantages.The improved AKAZE algorithm has a relatively small number of matches,the total matching time is low,and the real-time performance is better.while the improved BRISK algorithm has relatively weak real-time performance,but has a large number of matches.The image adaptability is good in complex areas.