Motion and structure factorization and segmentation of long multiple motion image sequences

Abstract

This paper presents a computer algorithm which, given a dense temporal sequence of intensity images of multiple moving objects, will separate the images into regions showing distinct objects, and for those objects which are rotating, will calculate the three-dimensional structure and motion. The method integrates the segmentation of trajectories into subsets corresponding to different objects with the determination of the motion and structure of the objects. Trajectories are partitioned into groups corresponding to the different objects by fitting the trajectories from each group to a hierarchy of increasingly complex motion models. This grouping algorithm uses an efficient motion estimation algorithm based on the factorization of a measurement matrix into motion and structure components. Experiments are reported using two real image sequences of 50 frames each to test the algorithm.