Opticalflow Estimation for Deformable Object Utilizing Superpixels and Motion Estimation Using Phase Correlation and CNN

Abstract

There are roughly two methods for estimating optical flow. One is the gradient method and the other is the block matching method. Each method has a few problems. In the gradient method, it is difficult to estimate optical flow when images include a noise and the difference in brightness between adjacent pixels is small. In the block matching method, an accuracy depends on the block division method and it isn't able to deal with a non-rigid body. Therefore, it is difficult to estimate the optical flow in a video that is easily affected by noise and exists deforming objects such as medical images. For each of these problems, we propose a method for estimating optical flow that is robust especially for object deformation while concerning a noise. The proposed method is based on the block matching method and has two improvements. One is to use Superpixels for block division. The other is to use the phase correlation for displacement estimation. Furthermore, we extend it by deep learning to be robust for a complicated change. The experimental results show the proposed method was robust to the object deforming in the medical image dataset.