Automatic panorama mosaicing with high distorted fisheye images

Automatic construction of a full view panorama is an active area of research in the fields of photogrammetry, computer vision and computer graphics. Using fisheye camera for panorama generation could greatly reduce the number of images thanks to its wide field of view. A key step in creating panorama with fisheye camera is to estimate the cameras' motion parameters which are usually estimated by established correspondences. However, establishing correspondences is a difficult thing particularly for high distorted images. For this reason, most of the traditional systems based on fisheye cameras for creating panorama are only designed for some specific configurations such that correspondences can be easily established. In this paper, we try to give a more general solution to establishing correspondences for a fisheye image pair. First, the feature detector unaffected by the non-linear distortion (like the MSER-regions) is used to extract features in original image. Then feature matching is realized on the virtual camera plane (VCP). Next the interested region is back-projected onto the VCP where it becomes a perspective one, so the standard affine invariant based matching algorithms can be employed directly. To combat data noise and improve the precision of the estimates of motion parameters, more correspondences are searched by applying guided matching strategy. Experiments showed that our algorithm can lead to the precise and reliable estimates and the final full view panorama is satisfactory.