Optical flow estimator with multichannel integration

In this paper, we propose a method of synthesizing several optical flow fields to form a unified one. The fields are first estimated from pre-filtered image with a series of band-pass filters. Our method is based on the fact that each of filters with different nominal spatial-frequency gives a specific region of greater determinant of Hessian matrix depending on the pattern. It is expected that integration of such regions will cover the whole image. This integration can be carried out by using maximum-likelihood method. Experimental results for both synthesized sinusoidal image and two kinds of captured image show that more stable optical flow field can be estimated by this multichannel integration.