Quasi-dense Matching By Rotation Model For Fisheye Images

The matching of image point correspondences is one basic problem in computer vision, which is also a core problem. What can be concluded is that, all integrated applications which involved two or multi-images, will be inevitably encountered the match problem of point correspondences between different images. So far almost researches about the matching of image point correspondences have been focused on normal perspective images. However, the fisheye images have more important application due to its extremely wide field of view, so it's rather valuable to search for the point correspondences between fisheye images.Quasi-dense matching is one classical method in matching the correspondent points, of which the choice of local transformation model is one important step. For the high distortions between fisheye images, this paper proposed a quasi-dense matching method based on local rotation model, and proved its better quality than some other methods. The main points in this paper are as follows:1. The performance characteristics of existing quasi-dense matching are researched deeply. And through the experiments it's found that the existing quasi-dense matching method based on the translation transform could not meet the distortion characteristics of fisheye image, while the quasi-dense matching based on affine transformation is not stable enough.2. The qualities of the existing local models are compared quantitatively through the simulation experiments, in which it shows that, during the situation with the noise, the rotation model and project model, affine model are comparable in accuracy, while the former one is more robust.3. For the distortions between fisheye images, a quasi-dense matching method based on rotation model for fisheye images is proposed. First, initial matching is obtained by using feature matching technique based on affine invariant, then quasi-dense match propagation is conducted around the initial seed points with disparity gradient limit, texture confidence measure and gray similarity constraint but no (epipolar) geometry constraint; after that, geometry constraint is estimated by using initial propagation matches with robust estimation strategy; constrained quasi-dense match propagation is done with the epipolar geometry constraint which leads to the more accurate quasi-dense corresponding points. Through the comparison of the methods based on ration model and classical affine model respectively, the results show that although our method's propagation precision in textured region is lower than the method based affine model, during the occluding and less textured areas, the more stable results can be obtained through our method.