A Research Of Image Feature Point Extraction Algorithms

As one of the most important local features in the field of computer vision,feature point in the image lays fundations in many visual applications.Because of the superior attributes including rotational invariance,scale invariance and so on,feature points are widely used in many applications,such as image stitching,image retrivial,target recognition and matching,3D reconstruction,visual reality and so on.Most of the traditional feature point detection methods are based on 2D grey level images.However,with the rapid development of image representation technique and image acquisition technique,more and more image formats appear,such as RGB-D images,3D images and panoramic images.Especially for the 3D images and panoramic images,the image data are quite different from the structure of 2D images,resulting that traditional 2D feature point detection methods is useless in other formats of images.Hence,how to design the feature point detection methods with high efficiency and rebustness becomes a hot topic in the field of computer vision.In this paper,we analyze various kinds of feature detection algorithms in terms of different image formats from 2D images to 3D images to panoramic images for the first time,and propose several robust 3D feature detection methods.Relative researches and academic contributions can be summarized as follows:(1)Summarize the feature point detection methods based on 2D images.In this paper,we divide traditional 2D feature point detection methods into three kinds: methods based on local gradient information,methods based on templates or learning and methods based on contour.Then we analyze the main principles,pros and cons of these algorithms and corresponding applications.(2)Introduce feature point detection algorithms based on other kinds of image formats.We discuss the feature point detection algorithms for 3D images,color space based,time and space based,RGB-D images and panoramic images.(3)Propose four original 3D feature point detection algorithms.On the basis of previous work,we propose four robust 3D feature point detection algorithms.They are(a)two kinds of 3D 3D feature point detection algorithms based on the catch points.(b)3D feature point detection algorithm based on the geometric measures and L0 sparse refinement of 3D mesh models.(c)3D feature point detection algorithm 3D feature point detection algorithm based on local and global information in the multi-scale space.Numerical experiments show that our proposed 3D feature point detection algorithms outperform other five state-of-the-art methods.