A kind of vision sensors consisting of an omnidirectional camera and a perspective camera is greatly useful. The omnidirectional camera can be used to monitor surroundings with wide-angle field of view and the perspective camera can be used to gaze at interest objects with high resolution. Image matching underlies many problems in computer vision. Many fairly mature image matching algorithms between perspective images have been proposed. However, because omnidirectional cameras have severe distortions, these previous image matching algorithms cannot be used for omnidirectional images directly. Researches on image matching involving omnidirectional images are still very few. In this dissertation, the main work is focused on the image matching between omnidirectional and perspective images, as well as their applications. The summary is as follows: ² A geometric invariant equation between omnidirectional and perspective images of the same scene consisting of planes is presented. This equation is based on cross ratios of pencils of lines and planes. It could be used to be a geometric constraint during image match propagation and to calibrate omnidirectional cameras. ² Based on the above derived invariant equation between omnidirectional and perspective images, a quasi-dense matching algorithm between such a pair of images is proposed. This algorithm does not require a camera calibration or a rectification for the omnidirectional image. First, the image matching algorithm between conventional wide baseline perspective images is used to obtain the initial matches between omnidirectional and perspective images. Second, a linear transformation is introduced to identify the area containing the corresponding point candidates. Then, the geometric invariant equation is computed as a constraint for quasi-dense matching. Finally, combining the computed geometric invariant with a best-first strategy of Lhuillier and Quan (2002), the quasi-dense point correspondences are calculated. The experiments with real data show that this algorithm could obtain quasi-dense point correspondences with even distribution which can satisfy the requirement of 3D reconstruction. ² The quasi-dense point correspondences obtained above are applied to 3D reconstruction. First, the geometric invariant equation and the computed point correspondences are used to calibrate the intrinsic parameters K of the omnidirectional camera under the division model of radial distortion. The...
修改评论