Research On The Mosaicing Algorithm For Large Area Aerial Images Captured By UAV

With the rapid development and popularization of UAV(Unmanned Aerial Vehicle)technology,UAVs are widely used in remote sensing in addition with their advantages.Working with large area aerial sequence images captured by UAV,this paper has done some embedded research on UAV images feature extraction and matching,distribution estimation and seamless blending based on methods of image processing and computer vision technology.Our main work and novelty are listed as follows:1.Aiming at the problem of matching multiple aerial images of large area,we proposed a matching strategy based on distribution estimation.After analyzing the imaging process of cameras on the UAV,the distribution map can be estimated by the feature matching between time-consecutive images in the sequence.The estimated distribution can be used as an inference for the matching of non-time-consecutive image pairs.Experiment results show that the number of useless attempt matching is largely reduced.2.For dealing with accumulative error and large algorithm complexity in the stage of global registration,we proposed a fast iteratively global registration method.This method has a very simple principle and avoids complex computation and solving for large matrixes like bundle adjustment.We applied all matched feature pairs to the iteration for more accurate registration parameters.Experiment results show that it is fast and performs well.3.To slove the problems of ?ghosting? and visible seam in blending,we combined Graph cuts and Multi-band blending to get a seamless pano.The optimal seam is searched according to Min Cut and Max Flow theory,and then Multi-band blending is performed.Experiment results show that a seamless mosaic of aerial image can be obtained by using our algorithm,and the overlapping region looks smoothly even with large explorsure differences.4.After implementation of our algorithm,a mosaicking experimental system was generated.Multi groups of data were used to test the system and our system works well in generating a seamless large resolution images.GPU was used in some models and the processing was largely accerated.