Study On UAV Image Matching Based On Affine Invariant Feature

The UAV remote sensing technology has become a hot topic which the countriesall over the world to reasearch, and has been widely used in the land resource survey,environment monitoring, weather monitoring and disaster emergency due to its fast andflexible, low costs, simple operation and maintenance and other advantages. However,the characteristics of small size and light weight result in differential rotation,translation, scaling, illumination, perspective and other aspects in same scene UAVacquisition images, and this increases the difficulty of UAV image matching. Therefore,grasping the characteristics of UAV image accurately, extracting features precisely andimproving the efficiency of features matching has become the the key of UAVphotogrammetry.In this paper, features extraction, matching accuracy and matching efficiency willdiscussed, using feature matching of UAV image as the object. On basis of summarizingthe characteristics of UAV image，features and matching algorithm, the robustness offeatures extraction and matching between point operators and scale invariant operatorhas been compared and verifying that the point and scale invariant operators aredifficult to satisfy the UAV matching problems with complex changes in illumination,scale, angle etc. Considering the irregular problems of the rotation angle, stripdeformation, forward and lateral overlap, feature extraction and matching’s relatedproblems based on affine invariant features have been studied. Using Hessian-Affineand MSER to extract features of common UAV image and tilt UAV images, the resultsshow that the affine invariant feature extraction operators are more effective to obtainhigh precision features. At the same time, according to the actual situation that higherdescriptor dimension, the principal component transform operator is introduced toimprove the traditional K-D tree matching strategy, so the PCA-KD tree algorithm isproposed. Experiments show that, compared with the traditional K-D tree, PCA-KD treeis complexity but it improves the feature matching error phenomenon effectively andimproves the matching precision.