T. Kuremoto, K. Koga, K. Kobayashi, and M. Obayashi (Japan)
optical flow, machine vision, template matching, aperture problem, interpolated quadratic surface, Beaudet operators
If the resolution of camera is not high enough, or the mo tion of object is slow sufficiently, than the estimation of velocity field (optical flow) must be performed on sub pixel level. The major previous techniques of optical flow computation can typically be divided into two branches: gradient-based methods and correlation-based methods (in volving frequency-domain, template matching, etc.). We propose a new technique here, considering both of the thoughts to realize a highly accurate estimation of slow op tical flow (all of velocities under 1.0 pixel/frame). The concept is to interpolate quadratic surfaces of brightness intensity in the vicinity of each pixel on successive frames, then under the rigidness assumption, the brightness and its gradient distribution of a local region on each frame are matched, the local velocity vector then can be calculated with an iteration algorithm. Experimental result shows the effectiveness of proposed method by using several syn thetic and real images.
Important Links:
Go Back
