The Research And Implementation Of Fast Panoramic Video Construction And Deghosting

The technology of panoramic image stitching has attracted much attention as theresearch focus of computer vision. A panoramic image usually refers to the image thatmeets normal view angle and binocular peripheral vision of human or360-degree fullrange of scene. Panoramic image stitching technology can be used in many fields suchas road monitoring, military infrared imaging, and minimally invasive surgery and so on.With the development of computer hardware, panoramic image stitching technology hasbeen widely applied to intelligent terminals and video surveillance.In the field of panoramic image stitching, the high computation complexity is amajor bottleneck restricting its application. This dissertation introduces a method of fastbuilding panoramic video and processing frames in real-time. The background and theresearch significance of image stitching are introduced initially in this dissertation. Thenthe paper also introduced the related works of image stitching and the main researchcontent of this dissertation. The general procedure of image stitching contains imagepre-processing, feature extraction and matching, homography estimation and imageblending. Since feature extraction is one of the most important steps of this work, thisdissertation introduced it in emphasis.SURF (Speed Up Robust Features) algorithm is applied to extract image features,and a new method is proposed to eliminate matching mistakes after feature extraction.Traditional transformation methods use only one homography matrix, so the accuracy oftransformation is low. In this paper, a new transformation method is proposed based onthe dual-homography to improve the stitching performance. Ghosting effect is auniversal problem in image stitching and difficult to avoid. Since human vision issensitive to this phenomenon, it is important to remove it. The ghosting can beclassified into two categories by the cause of formation. One is caused by the movingobject going through the overlapping area. Since the parallax of cameras, the movingobject will generate the ghosting phenomenon in the overlapping region. The other iscaused by the seam line especially in video application because the seam line is theextrema of parallax. This paper combines image subtraction and smooth transition blending methodto eliminate ghosting effect. The deghosting method of this dissertation is comparedwith simple linear blending, Poisson blending and multi-band blending. Theexperimental result reveals that the method in this paper has better performance than themethods mentioned above both in effectiveness and time complexity. The method thispaper proposed can be implemented in general personal computers. The experimentsreveals the frame rate of this method can reach to30frames per second and achievereal-time requirement.