Research On Method Of Aerial Image Mosaic In Building Areas

UAV-based reconnaissance is an important means of security reconnaissance in all countries in the world today.Low-altitude reconnaissance in building areas is an important branch of security reconnaissance.By stitching low-altitude aerial images,high-definition panorama with rich information can be obtained,which effectively improve the efficiency of investigation.However,due to the large parallax caused by the high buildings in the image,the existing stitching algorithms are often accompanied by problems such as misalignment and ghosting in the stitching of low-altitude aerial images in building areas.In response to the above problems,this paper proposes two effective low-altitude aerial image stitching methods for building areas based on the segmentation results of high-building areas: image stitching methods based on large parallax complementation and image stitching methods based on hybrid transformation model.The specific work is as follows:First,based on the SegNet network,an improved semantic segmentation network SegNet＿UAV is proposed to realize the segmentation of high buildings in aerial images.Improvement methods include: increasing the resolution of the input layer to reduce the loss of image information;this paper proposes the Encoder Unit and the Decoder Unit structure introducing two-way feature extraction in the unit to enrich feature extraction methods;adding residual connections in the unit to reduce the gradient disappearance during network propagation.Secondly,an image stitching method based on large parallax completion is proposed.This method first segmented the high-building area,and then stitched the other areas except the high-building as a background image to eliminate the interference of large parallax on the image,and then select the best orthophoto image from sequences images of the high-building area,and then patch it to the stitching result of the background images to get the final stitching result.Finally,an image stitching method based on the hybrid transformation model is proposed.This method seeks the optimal global transformation based on avoiding high-building areas in the overlapping area as much as possible.After obtaining the global optimal transformation,combined with the local transformation,the local alignment of the overlapping area is more accurate and the non-overlapping area is as undistorted as possible.In order to verify the effectiveness of the method proposed in this paper,a large number of experiments have been carried out in this paper.The experiments show that when dealing with the problem of low-altitude aerial image stitching in construction areas,the method proposed in this paper has better stitching results than the panorama obtained by classic algorithms such as APAP.