Multi-resolutional optical flow estimation with local optimization

In this paper, multi-resolutional optical flow estimation using a gradient-based method with local optimization is proposed. The method computes image motions in a fine resolution at points where image displacements are small, and in a coarse resolution at points where image displacements are large. Considering the effect of a low-pass filter that is applied to make the coarse resolutional image, some parameters are set for each resolution as follows: (1) size of a local window where optical flow is assumed to be constant is enlarged as resolution is decreased; (2) threshold of the smaller eigenvalue of the Hessian of image is determined by the power of white noise, impulse response of the low-pass filter and derivative operator. The proposed method is tested and discussed on synthetic and real image sequences. We find an optimal resolution for velocity to estimate optical flow in sequences where white noise is added.