Research Of Fisheye Image Keypoint Extraction And Matching Based On Spherical Camera Model

Fisheye lens has a wide field of view(180°), fewer images can cover the whole three-dimensional space. So it is widely used in visual surveillance, robot navigation, virtual reality and other fields. These applications are based on the image feature extraction and matching. The matching accuracy of the feature points and the number correspondent points directly affect the quality of these applications. However fisheye image having a large distortion, especially in the wide baseline situation, the perspective projection image feature point extraction and matching algorithm for the perspective projection is not able to meet the requirements.This thesis focuses on the characteristics of distortion of fisheye image. To remove the distortion, a deep research is taken on the fisheye lens image model and the algorithm of image feature extraction and matching. A model of fisheye lens based on the scale changing and homogeneous coordinates is designed. Moreover a scale space theory for fisheye image based on sphere Gaussian and the matching and removing wrong matching algorithm based on scale changing is designed.Firsty, a review of keypoint extraction and description algorithm is taken. There is no suitable algorithm for fisheye image. Then the algorithm for pinhole image directly applied to image with distortion is discussed. Analyzing the reason why it is not suitable for distortion image. To remove the affection of distortion, a fusion of image model and traditional algorithm is designed.Secondly, this thesis studies the image model of fisheye lens. The scale disunity in one image is found. So a image model for fisheye lens is designed which is based on the fusion of sacle changing and homogeneous coordinates. A calibration method is proposed combined with Zhang Zhengyou calibration. The test of verification is carried. It is the basis of improvement of the algorithm in this thesis.Thirdly, this thesis studies the algorithm of feature extraction. Through the relationship between the scale space and the heat conduction in image, combined with the image model the spherical Gaussion and scale space theory is deduced. A convolution method on original image is designed to eliminate the interpolation, which is an adaptive filter kernel convolute with the image. Finally the DoG space is built and keypoint is extracted.Finally, this thesis designed the descriptor for the keypoints. At the beginning D-Nets algorithm is improved to illustrate the necessarity of the fusion of image model with traditional descriptor. Then the modification is carried on the region based algorithm SIFT. The correction of gradient is designed in the modification. At last, a wrong matching removing algorithm is designed to improve the correct rate.