F. Wenzel and R.-R. Grigat (Germany)
3D Computer Vision, Structure from Motion, Hough Transform, Motion Discontinuities, Motion Parallax
Epipoles, which are the mutual projections of camera centers and determine the direction of heading in the case of a moving camera, can be estimated in many ways. Some approaches are based on the epipolar ge ometry of a scene and require, or are part of, a calibra tion process. Others analyze the dense motion field, also known as the optical flow. We follow the second approach and assume that the optical flow has already been computed. Estima tion of the epipole is done in two steps. First, we identify motion discontinuities in the flow field by fil tering the optical flow. This way, information about motion parallax can be extracted. We introduce dif ferential homographies as a means to justify our ap proach. In the second step we use this information for a Hough transform that reveals epipoles in param eter space. We choose a two-dimensional mapping of spherical coordinates in order to deal with both finite and infinite epipoles and present experimental results with real and synthetic data.
Important Links:
Go Back