New optical flow approach for motion segmentation based on gamma distribution

Abstract

This paper provides a new motion segmentation algorithm in image sequences based on gamma distribution. Conventional methods use a Gaussian mixture model (GMM) for motion segmentation. They also assume that the number of probability density function (PDF) of velocity vector's magnitude or pixel difference values is two. Therefore, they have poor performance in motion segmentation when the number of PDF is more than three. We propose a new and accurate motion segmentation method based on the gamma distribution of the velocity vector's magnitude. The proposed motion segmentation algorithm consists of pixel labeling and motion segmentation steps. In the pixel labeling step, we assign a label to each pixel according to the magnitude of velocity vector by optical flow analysis. In the motion segmentation step, we use energy minimization method based on a Markov random field (MRF) for noise reduction. Experimental results show that our proposed method can provide fine motion segmentation results compared with the conventional methods. © 2010 Springer-Verlag Berlin Heidelberg.