ROBUST POINT CORRESPONDENCE AND POSE ESTIMATION

摘要

Feature correspondence between two or more images is crucial for many image analysis tasks. Recently, there has been a boost of interest in the correspondence estimation problem due to the development of the technique of 3D camera pose estimation, which is called Structure From Motion (SFM), A good initial set of feature correspondence is required to obtain the camera motion parameters. This paper proposes a new robust point correspondence algorithm for 3D object recognition based on the corner points of 2D projective image and template. The point correspondence method based on Singular Value Decomposition (SVD) has been extended in order to cope with large rotation, translation and scaling. The pose estimation algorithm is based on the geometric computation for simulated camera motion, which is known as the optimal SFM. Our proposed algorithm solves the problem of SFM that it can handle two sets of features with different number of points. The simulation results from a number of alphabetic letter images show that our algorithm is able to deal with large rotation, translation and scaling. The experiments also show a good result for 3D plane reconstruction using our algorithm.