Human posture recognition: a proposal for mean eigenspace

We address a mean eigenspace (MES) approach in this paper, which overcomes the limitations of conventional eigenspace technique for human posture or flexible objects recognition. A MES is produced by taking an average of some selected eigenspaces. In fact, a mean of similar eigenspaces of different human models creates an optimized visual appearance, and unknown postures are recognized comparing with it. We do not need the unlimited number of eigenspaces for producing a MES. The present study proposes an idea for the appropriate eigenspaces selection. This study also employs edge images for creating an eigenspace in order to minimize the dress effect. We have conducted experiments employing thirty subjects wearing various clothes (including different sexes, races and nationalities). Experimental results show robustness and effectiveness of the proposed method.