Research On Key Technologies Of UAV Image Stitching

Remote sensing satellites has the ability observe the earth macroscopically andobtain a wide range of remote sensing images. But it is unable to provide remotesensing images in the first time for emergencies because of the limitation of satellitereturn cycles. Compared with satellite, UAV images can be obtained more flexiblyand timely in emergencies. However, UAV images are generally with small breadthand high overlap. Thus, hundreds or thousands of images are need to be stitchedtogether in order to obtain the whole view of target area. According to thecharacteristics of UAV image stitching, this paper presents a fast UAV imagestitching method and the main work of this paper is as follows:(1) Compare The advantages and disadvantages of the SIFT, SURF and ASIFTin the similar resolution image (UAV images) matching and in different resolutionimages (UAV images and Google Map images) matching. On this basis, SURF isconsidered to be more suitable for matching of UAV images matching, and ASIFT ismore suitable for matching of UAV images and Google Map images.(2) The mismatched points cannot removed completely using the RANSACalgorithm, and the matched points are tend to be clustered. To deal with this problem,a feature point matching method constrained by search range and projective invariantpolygon is proposed. Firstly, using the ANMS in feature point extraction to make thenumber and spatial distribution of the feature points be more reasonable. Then, thelarge scale feature point matching result is used to constrain the following matchingprocesses, and the mismatching points are removed in each stage of feature pointmatching using the similarity of a planer polygon and PEPP its (perspectiveequivalent planar polygon). The proposed method can not only ensure the accuracy offeature points matching, bit also can remove feature points located on tall objects, andthus avoid those feature points misleading the calculation of the transformation modelbetween images.(3) For the questions of error accumulation in multiple image stitching and fastdetermination of geographic location and orientation of the stitched image, ageographic constrained UAV image global stitching method is presented. Firstly,cropping operation is perform for UAV images based on their calculated degree ofoverlap. Then, geographic location and orientation are determined by performingregistration between UAV image and Google Map images. Finally, all the involved UAV images are stitched and fused based on their geographic information. Thismethod can ensure the registration error of each UAV image to be independent andavoid error accumulation. The final stitched image have a good visual effect and thegeographic information can be quickly obtained.The proposed method was validated on the UAV images in South LakeCampus of China University of Mining and Technology.