Font Size: a A A

Research On Feature Correspondences For Fisheye Images

Posted on:2011-08-23Degree:MasterType:Thesis
Country:ChinaCandidate:Y P TianFull Text:PDF
GTID:2178360305495366Subject:Pattern Recognition and Intelligent Systems
Abstract/Summary:PDF Full Text Request
The matching of image point correspondences is to search for the point pair which is corresponding to the same spatial point between the different images. For the almost computer vision problems (such as motion estimation, object recognition, target detection, and 3D reconstruction), establishing the point correspondences is the most basic step, so of which is also a core problem. So far lots of point corresponding matching algorithms can be obtained from the inland and abroad, however most of them have been focused on the perspective images. The fisheye images have important application due to its extremely wide field of view. However, it's more difficult to matching point correspondences due to its severe non-linear distortion. In this paper, the main work is focused on how to match the corresponding points between fisheye images, and some conclusions are obtained through our work. The main points are as follows:1. The influence about robust estimation strategy to linear parameters estimation is researched and a practical robust estimate approach is proposed. Although both the distortion and the data noise of the region which is far from the image center are more severe than the region which is close to the center, the data of the region far from the center are quite important to estimate the parameters. Based on the above observation, a reasonable assumption is made that different regions have different data noise, and we give a strategy to gradually select more and more inliers from a noise corrupted data set. Some performances can be shown by using this strategy. One is that more data points can be established, which are quite useful to the linear estimate problem from a noise-corrupted data set. The other is that data set contains more data points which are far from the image center. The above two performances can make both the precision and stability of the final motion parameter estimation improved greatly.2. Three typical distinctive point detector based on contour are implemented by program. Analysis and comparison have also been done to these algorithms, which are extremal curvature point detector, contour flexibility minimum point detector and multi-scale visual curvature point detector. By comparison from detecting accurate, parameter setting and the algorithm implement efficiency, it shows that multi-scale visual curvature point detector performs better than the other detectors.3. A frame of matching correspondences for fisheye images is proposed, and an implementary approach is given. Due to the severed distortion of the fisheye images, few correspondences will be gained from the conventional affine invariant methods. So these methods can not meet the request of practical application. The main idea of matching frame we proposed is as following. First, initial matching is obtained by using based affine invariant technique, and rough geometry constrain model is estimated. Though this model is not very precise, it can be used to guide continuous process to establish more corresponding matching. Then constrain model is updated gradually. Finally, we obtain final stable result. Experiment results show that we can establish a large number of reliable and stable distinctive point correspondences, which make the precision of motion estimation and geometry estimation improved greatly.
Keywords/Search Tags:Fisheye Image, Feature Matching, Motion Estimation, Contour Feature, RANSAC
PDF Full Text Request
Related items