Research On Feature Correspondences For Fisheye Images Based On Dynamic Programming

Feature matching is not only a fundamental but also a difficult problem in computer vision. Compared with the traditional perspective images, fish-eye images have more serious non-linear distortion and the illumination changing in different regions, which make feature extraction and matching more difficult. The traditional methods on image matching are most based on textural features which have a thin skin to varying illumination. Methods based on shape features are steady to varying illumination, so these means are hot spots of image matching. The research on feature correspondences for fisheye images is combined with dynamic programming and some fruit has been achived. The main points are as follows:1. An novel shape descriptor of planar contours is proposed in this paper, called contour flexibility. First, maximally stable extremal regions was distilled, then a sequence of landmarks can be easily obtained from a simple closed plannar by uniform or nonuniform sampling. Then, a contour flexibility sequance can also be computed by the arithmetic which is expounded in this paper. Experiment results show that the contour flexibility has the capability to reflect the feature of the contour, and this method is reliable and stable.2. Performance test with typical date is implemented, and different parameter has also been considered. Experiment results show that this method is stable to translation, rotation, and the size scale.3. With the flexicibility sequences and dynamic programming, methods focusing on local features is implemented, which is adaptable to local distortion. However, the methods focusing on local features may fail to match objects of the same class when the contours have significant deformation or noise. We just develop a method in favour of the global shapes of objects with the contour flexicibility which is stable to noise. Our scheme for matching two shapes with the local and global features is combining the local and global method, which is stable to noise and distortion. Experiment results show that this method is feasible and complementary to methods based on textural features.