Research And Application About Small Unmanned Aerial Vehicle Low Altitude Remote Sensing Image Registration Based On Local Outlier Factor And Rotation Invariant Feature

In the field of ground monitoring,the small unmanned aerial vehicle(SUAV)is widely used due to its convenience and low cost.However,due to the influences of natural factors,human factors,and equipment factors,low-altitude remote sensing images acquired by SUAVs in the same scene will have rotation,scale changes,nonrigid distortion,and the mixture of the above.These issues bring problems to subsequent applications.Therefore,it is necessary to use image registration as a preprocessing method to align the images acquired at the same place in different time periods to facilitate subsequent applications in the fields of natural disaster assessment,environmental monitoring,and change detection.Aiming at the characteristics of low-altitude remote sensing images taken by SUAV,such as low overlap,high image resolution,and large distortion in the image,a feature-based image registration method is designed in this paper.In this method,a rotation invariant feature is designed and the feature matching problem is transformed into an outlier elimination problem.The local outlier factor is used to find out the false samples to achieve the purpose of mismatch removal.The main contributions of this article are:(1)The neighborhood information in the image is deeply explored,and the rotation invariant feature is designed.The rotation invariance feature is invariant to various problems in low-altitude remote sensing images of SUAV,such as rotation,scale changes,and non-rigid distortion.(2)The neural network is used to fuse the neighborhood structure information with the image information,which enhances the ability of the designed rotation invariant features.(3)Using spatial position information and motion consistency information to construct samples for local outlier factor,and designing an adaptive weight adjustment strategy in the sample distance calculation step to weaken the value of coordinates and enhance the value of consistency of motion when the feature points are close.(4)In order to enhance the algorithm’s robustness under high mismatch rate,the result of the fusion of neighborhood structure information and image information in(2)is used as a penalty factor to penalize the sample distance.(5)In order to increase the recall rate of the algorithm and enhance the discrimination between mis-matches and true matches,a step to reduce the spatial difference is designed to perform rough registration on the feature point set.The proposed method can be applied to a variety of ground monitoring fields as a pre-processing method.This article introduces the application of ground garbage monitoring combined with the proposed image registration method,change detection and neural network technology.In the experimental part,the proposed method in this paper is compared with other six representative methods for feature matching and image registration.Experimental results show that the proposed method can well balance precision and recall,and achieves the best comprehensive performance compared with other methods.