In this paper, for the scene of panoramic driving recorder, we obtain a panoramic video using two video streamings taken by an opposite pair of fixed binocular fisheye cameras. We calibrate the parameters of fisheye camera and correct distorted images using a cylindrical mapping model of perspective projection. Then we extract the SIFT features for two frames of corrected image and remove mismatching points as well as figure out a homography matrix using the RANSAC algorithm. Finally, we use the linear weighted fusion method between the two frames of cylindrical image to get a panoramic image. In this experiment, we use cylindrical projection which is consistent with human visual characteristics. And comparing with perspective projection, cylindrical projection reserves more information of view which can help drivers to observe more important information.