Multibody motion segmentation for an arbitrary number of independent motions

Abstract

We propose a new method for segmenting feature point trajectories tracked through a video sequence without assuming a number of independent motions. Our method realizes motion segmentation of feature point trajectories by hierarchically separating the trajectories into two affine spaces in a situation that we do not know the number of independently moving objects. We judge that input trajectories should be separated by comparing the likelihoods computed from those trajectories before/after separation. We also consider integration of the resulting separated trajectories for avoiding too much segmentations. By using real video images, we confirmed the efficiency of our proposed method.