Regularized Pel-Recursive Motion Estimation Using Generalized Cross-Validation and Spatial Adaptation

Vania EstrelaRicardo Tadeu Lopes

This paper addresses the ill-posed problem of estimating the displacement vector field (DVF) resulting from the optical flow via picture element recursive (pel-recursive) algorithms. The term pel-recursive comes from the fact that the first proposed algorithms used the motion vector of the previous pixel to generate the initial estimate for the motion of the current pixel. The problem of estimating the motion vector at each pixel location, in a pel-recursive manner, is formulated as a regularized least-squares (RLS) problem. A data-driven method to regularize adaptively the motion vector estimates on a pixel-by-pixel basis is proposed: the Generalized Cross-Validation (GCV) approach. Experimental results are presented with both noisy and noise-free images which demonstrate that the proposed regularized motion estimates outperform significantly previous motion estimating strategies which do not regularize the DVF on a pixel-by-pixel basis. The main contributions of this work are: (a) a method to estimate the regularization parameters, which can also be viewed as the statistics of the motion field and the noise, adaptively based on local properties of the image; (b) a discussion demonstrating the advantages of regularization in the context of our problem; and c) the introduction of the idea of spatially adaptive support. According to this idea, there is an optimum neighborhood shape around each pixel that is assumed to have the same motion as the current pixel. This allows accurate motion field calculation in areas of motion discontinuities.

Caso o link acima esteja inválido, faça uma busca pelo texto completo na Web: Buscar na Web

Biblioteca Digital Brasileira de Computação - Contato:
     Mantida por: