Image Stitching Technology For Large-scale And Multi-view UAV Remote Sensing Images

The unmanned aerial vehicle(UAV) has the advantage of high flexibility and efficiency, low cost, operating conveniently, being highly targeted. Therefore, UAVbased remote sensing imaging has been widely applied for geological testing, disaster monitoring, urban planning and military applications. However, since the UAV remote sensing images have the characteristics of high resolution, changeable perspective, and huge quantity of images, rapid and accurate automatic stitching of these images is still a challenge. This paper targeted at the UAV-based remote sensing images, which are with no accurate georeference information and camera parameters. Two image mosaic algorithms for UAV remote sensing image were studied, in order to achieve global optimization and ensure mosaic quality.Current registration algorithms for remote sensing images work well when georeference is available, which is usually not the case for UAV-based image. Many factors affecting the registration results for UAV-based images: comparing to traditional aerial photography, UAV-base imaging encounters quick changes in orientation due to the influences from air flow because the UAV itself has a relatively light weight; the captured images arrange unevenly and angular variation of each two adjacent images could be very large; information on the pose of the UAV for each image may be unavailable. All these cause great difficulties for the image mosaic for UAV-based remote sensing images, especially when the number of images are huge.The main work presented in this paper is as follows:(1) On the basis of the traditional image stitching algorithm based on SIFT, this paper proposed a switching algorithm for remote sensing images with global optimization on feature points and stitching sequence selection. Two stages were applied in order to improve the results: in the feature points matching stage, the KDtree approach was utilized instead of the traditional LSH method; Besides Euclidean distance ratio and median filter were combined to select the feature points from coarse to fine-grained matching matching, thus greatly improve the speed of matching feature points; in stitching sequence selection stage, the best reference image was identified based on the image overlay density, so that the image stitching error of multiple images after registration is minimized.(2) Considering that the reference image selected above does not necessarily provide the smallest mosaicking error, this paper proposed another image mosaicking algorithm based on 3D point cloud. Firstly, sparse point cloud for the scene was restored based on SFM; Then the best fitting was achieved by minimizing the distances of the feature points on all images to the reference plane; Finally, the corresponding relation between the images was established according to the recovered feature, so that the panoramic image could be mosaicked for the case that the camera pose was unknown.(3) Two groups of UAV-based remote sensing images captured from Xixi Wetland were used to evaluate the results of the two algorithms proposed. The final results showed that the first method proposed in this paper can improve the registration speed while ensuring the accuracy; and can be effectively used for, and the effect of the second method meets the requirement of the practical application for UAV image stitching with unknown camera pose.