Soccer Player’s Pose Recognition by Creative Search for Generating Free Viewpoint Images

Abstract

The low-cost generation we have proposed for a free viewpoint image of soccer games uses a multiple viewpoint image database designed with computer graphics for recognizing player poses in the real image and for generating virtual scenes. A pose is recognized from a player’s silhouette and applying parametric eigenspace method. Because only silhouette information is used, however, limb positioning may be backward or the body misdirected. Our new proposal for excluding misplaced left and right limb poses assumes that changes in a pose, especially limb positioning, between sequential frames are continuous, so the limb positioning in 3D space and 2D images can be determined and the search range restricted in eigenspace. We also propose generating continuous frames for cases in which a correct pose exists outside of the restricted range by setting an initial state and handling error.