A neural-based method for optical flow estimation using phase correlation

Abstract

Motion estimation for image sequences is one of the most important tasks in computer vision. Thus, many methods have been proposed to solve this problem, but even so, it still lacks a generic method that determines motion in all situations and for all types of objects. In this work, we propose a two-phase connectionist neural method for motion estimation in the frequency domain that takes discontinuities into account. In the first phase, the most probable motion of each pixel is estimated using self-organising maps principles and the phase correlation method. The second phase consists in regularising the displacement field that considers the discontinuities. When tested and compared with other approaches on both synthetic and real image sequences, our method showed good performances according to the following criteria: precision, regularity, resistance to noise and running-time. Moreover, it could estimate the motions in cases where rotation and scaling are required.