Research Of Video Stitching Technology Based On Multiply Wide-angle Cameras

Currently, the video mosaic is a technology that stitches the videos with overlapping perspectives to form wide field of video image. Compared with monocular video image, it can provide more information while presenting greater size of video pictures. With the constant popularization and inexpensiveness of multimedia mobile equipment, video mosaic technology has apparently become a research hotspot in image processing field. Compared with image mosaic because of real-timeness and synchronism of video, video mosaic is more complicated. This paper used three-channel video to conduct experiments, and stitched it based on improved SIFT algorithm to research the video mosaic technology in this paper.The main contents of this paper are as follows:(1) This paper firstly introduced the theories related to image mosaic, including imaging model of camera, relationship between coordinate transformations of images as well as motion model of camera and the meaning of image registration. Also, this paper simply introduced common registration technologies, which laid a foundation for subsequent part of this paper.(2) To balance the relationship between the quantity of cameras and the market, this paper selected wide-angle cameras to form images. As the image edges captured from a wide-angle camera are seriously distorted, so the distortion must be disposed at first. On the basis of selecting a calibration method based on linear feature, this paper also put forward an indicator function of bending measure function to provide different weight factors for the curves at different distances close to the image center.(3) SIFT algorithm was analyzed in details, and was improved in line with the real-timeness requirement of video mosaic. Firstly, the area of interest of the images was limited. SIFT algorithm extracted over many and dense feature points, which was not conducive to subsequent matching of feature points and calculation of projection transformation matrix. Therefore, it was necessary to improve it. Firstly, the searching range of extreme points was expanded. Secondly, the distance of the feature points extracted were restrained. The results of simulation experiment showed that improved SIFT algorithm was faster than the original algorithm, and the feature points distributed more uniformly; meanwhile, it also maintained sound robustness of this algorithm for illumination and rotation.(4) This paper selected matching algorithm of feature points based on the ratio of the closest neighbor and the next-closest neighbordistances to conduct coarse matching for the feature points. Also, this paper selected BBF search strategy to search closest neighbor and the next-closest neighbordistances, purified the feature points by means of RANSAC algorithm, and then estimated the projection transformation matrix by using the purified feature points. To remove the mosaic joints after registering the images, this paper firstly analyzed existing image fusion algorithms and compared them through experiments. Weighted average fusion algorithm could retain the details of the images satisfactorily and acquire better quality of fused images. That’s why this algorithm was selected to fuse images in this paper.(5) In the imaging process,because of the position of the camera and the parameters of the camera are fixed,we just only need to extract the image feature point of the first frame to calculate the projection transformation matrix,the rest frame is operated by using this projection transformation matrix.And the OpenMP parallel programming method is used to optimize the whole system,to improve the program efficiency.In order to keep the synchronous acquisition of video image, we use DirectShow technology to obtain the video image.At the same time,a kind of storage mechanism is designed to develop a buffer for each video stream.