Representing Human Motions by an Eigenspace

概要

論文の詳細を見る
This paper describes a technique for representing human motions employing an eigenspace technique. A video image sequence obtained from a human motion defines a multi-dimensional eigenspace in which the image sequence is represented as a curved line. The trajectory of this line should depend only on the motion and not on a person changing his/her dress in order to proceed to the recognition of unknown human motions. A different curved line is, however, created each time a person wears different dresses, since the original eigenspace technique employs initial gray-level images. To escape from this difficulty, the present technique employs blurred edge images of the original. In the performed experiment, 5 persons wearing respective clothes play 6 actions of a cricket umpire. A 6-dimensional eigenspace is created from their blurred edge images, in which one is able to see that the trajectories of the obtained 6 curved lines depend on the actions and not on the persons.