View-invariant pose recognition using multilinear analysis and the universum

Abstract

This paper presents an approach to full-body human pose recognition. Inputs to the proposed approach are pairs of silhouette images obtained from wide baseline binocular cameras. Through multilinear analysis, low dimensional view-invariant pose coefficient vectors can be extracted from these stereo silhouette pairs. Taking these pose coefficient vectors as features, the Universum method is trained and used for pose recognition. Experiment results obtained using real image data showed the efficacy of the proposed approach. © 2008 Springer Berlin Heidelberg.