| Drone panoramic aerial images meet the pursuit of wide view field and high-resolution,which is closely relevant with the people’s lives and work.Therefore,it has been widely used in post-disaster rescue,mineral exploration,field forensics and other fields.In general,a drone needs to fly at a low altitude to guarantee the high resolution of images.However,it results in small view for a single image and fails to provide sufficient useful information in this case.On the other hand,there would be large distortion for a wide-angle lens,which is not suitable for technical fields that require high image realism.Therefore,when adopting ordinary lenses in aerial photography,there is a trade-off between view field and definition.Thus the research on panoramic stitching of aerial images is particularly important.However,aerial image is sensitive to light,rotation and scale variance.It requires better matching performance and image fusion effect for stitching algorithms,as well as maintaining a certain calculation speed.Based on the above analysis,this thesis conducts a comprehensive study on the issues of improving the speed and accuracy of image registration,and the effect of image fusion.The main work and innovations are as follows:(1)Aerial images preprocessing.Aiming at the noise generated in taking and transmission of aerial images,four denoising algorithms are introduced in experiments.Qualitative results show that the bilateral denoising algorithm performs best.In addition,since input images would be blurred due to the influence of air quality and other factors during taking,a histogram enhancement method is adopted in the preprocessing stage to increase the detailed information of an image.Moreover,aiming to the geometric distortion by the nonlinearity system,the camera is calibrated by Zhang’s calibration method.(2)Feature extraction and description.In order to output real-time panoramic aerial images on a drone,the feature extraction algorithm should ensure high speed performance.In this thesis,four sets of experiments are carried out on SIFT,SURF,ORB,and BRISK.Compared with other two features,ORB and BRISK with better efficiency are selected for further research.(3)An improved image registration method.For the problem that large number of iterations is needed in the RANSAC algorithm,it is accelerated by a method of dividing the matching feature set in half,terminating the error fitting process in advance and uniform partition extraction.Combining the idea of image segmentation and bilateral matching strategy,an improved image registration framework based on GMS-RANSAC is proposed.By comparing this framework with ORB + RANSAC,BRISK + RANSAC,NNDR + LSH,GMS and Improved GMS algorithms on multiple sets of aerial images,it verifies the advantages of matching speed and matching accuracy in the proposed framework.(4)An improved fusion algorithm for serial aerial images stitching.For the problem that the trigonometric function fusion algorithm can only eliminate stitching seams in one direction,a two-dimensional trigonometric function fusion algorithm is proposed,which is based on an adaptive weighted normalized smoothing function.In order to verify the fusion effect of the improved fusion algorithm,four methods of fade-in and fade-out fusion method,trigonometric function weight fusion method,direct average fusion method and the improved algorithm are compared and analyzed by multiple experiments.The results show that the improved algorithm has outstanding fusion performance,whose advantage is more obvious when the image exposure difference is relatively large. |
