Research On Aerial Image Stitching Technologies Based On Features

With the development and popularization of science and technology, people become increasingly dependent on the images as an information tool, and the requirements for the application of the image processing become increasingly high. Image stitching technology can well meet the demand for large image with high resolution, wide viewing angle. In this paper, we do the research on feature-based aerial image stitching.Image registration based on feature extraction is the key of image stitching. In this paper, a simulation comparison is made against the most popular feature extraction algorithms. The feature extraction algorithm is mainly divided into two steps: feature detection and feature description. The feature detection algorithm includes SIFT and SURF based on blob-like response filters, and imporved multi scale FAST corner detector. The descriptor includes the SIFT and SURF based on gradient information, and other binary descriptors like BRISK and FREAK based on the comparison of pixel values. Based on the characteristics and project demands in aerial images, this paper proposed a new aerial image stitching algorithm based on SURF and FREAK, which can guarantee to extract a sufficient number of features, but also improves the computational speed. Experimental results show that it can get good results if there is no large parallax in the scene.Due to the condition of aerial images, when there is a larger parallax in the scene, it will produce large errors. So this paper introduces two improved methods, the first proposes as-projective-as-possible warps based on Moving DLT, in view of the traditional single global projective model. The as-projective-as-possible warps can not only keep the global projection, but also ensure the accuracy of local registration. The second method is seam cutting. This paper proposes an algorithm based on improved RANSAC and seam cutting, which can handle large parallax more efficiently. When stitching multiple images, there will be cumulative errors. So we realize image stitching of multiple aerial images through selecting the reference frame and bundle adjustment optimization method. At the end, the paper realized multi frame aerial image stitching system based on the platform of DSP and PC, and both achieved good effect of stitching.