Panoramic video stitching of dual cameras based on spatio-temporal seam optimization

This paper proposes a panoramic video stitching algorithm based on seam optimization, which aims at stitching two videos taken by two wide-angle cameras into a single 720-degree video. The use of only two cameras makes the parallax of the dual videos very large, while previous stitching methods based on deformation or seams incur problems like distortion, blur and ghost. To solve these problems, we improve the graph-cut algorithm to compute the optimal seams in the overlapped regions. For the spatial and temporal consistency of the panoramic video, foreground detection and Gaussian filter are employed to generate a sequence of smooth seams. Besides, a quantitative evaluation on the seam quality is proposed for the linear fusion of the stitched frames. Compared with previous methods, our work can effectively reduce the distortion, blur and ghost artifacts, as well as maintain good spatial and temporal consistency of the panoramic video as evidenced by the experiments.