Accurate Dense Optical Flow Estimation Using

An accurate optical flow estimation algorithm is proposed in this paper. By combining the three-dimensional (3-D) structure tensor with a parametric flow model, the optical flow estimation problem is converted to a generalized eigenvalue problem. The optical flow can be accurately estimated from the generalized eigenvectors. The confidence measure derived from the generalized eigenvalues is used to adaptively adjust the coherent motion region to further improve the accuracy. Experi- ments using both synthetic sequences with ground truth and real sequences illustrate our method. Comparisons with classical and recently published methods are also given to demonstrate the accuracy of our algorithm. technique by using the affine motion model. He defined the tensor by projecting the image onto a second-degree polyno- mial and integrating the affine model into the tensor. The affine parameters were solved as a linear system. Based on an idea used in (4), we derive the 3-D structure tensor technique in a different way. We show that when we use a parametric motion model to construct the 3-D structure tensor, the motion parameters and subsequently the optical flow can be found using generalized eigenvalue analysiswithout actually solving a linear system. This is demonstrated by using an affine motion model as an example. Since a reliable confidence mea- sure can be derived from the generalized eigenvalues, it can be used to dynamically adjust the neighborhood to include a wider area of coherent motion, so that the estimated flow is more ac- curate and robust to the aperture problem. Here the coherence is in terms of the parametric motion model instead of the optical flow. The rest of this paper is organized as follows. Section II for- mulates the optical flow estimation problem as a generalized eigenvalue problem. Section III analyzes the relationship be- tween generalized eigenvalues/eigenvectors and optical/normal flow. A confidence measure, which is then used to guide the neighborhood adjustment, is also defined in terms of the gen- eralized eigenvalues. Experimental results using synthetic and real image sequences are provided in Section IV. The results are compared to a few classical methods and some recently pub- lished methods. Section V gives conclusions.