HUMAN POSTURE RECOGNITION: EIGENSPACE TUNING BY A MEAN EIGENSPACE

This paper investigates an appearance-change issue due to various human body shapes in an eigenspace analysis, which is responsible for generating person-based eigenspaces employing a conventional eigenspace method. We call this a figure effect in this study for this phenomenon. As a consequence, an appearance-based eigenspace method cannot be effective for recognizing human postures with its present available formulation. We propose to employ a generalized eigenspace for avoiding this problem, which is developed by calculating a mean of some selected eigenspaces. We also investigate a dress effect due to human wearing clothes in this paper. The study proposes image pre-processing by Laplacian of Gaussian (LoG) filter for reducing the dress problem. Since the proposed method tunes a conventional eigenspace as an appropriate method for human posture recognition, the proposed scheme is known as an eigenspace tuning. An eigenspace called a tuned eigenspace is obtained from this tuning scheme and it is used for further recognition of unfamiliar postures. We have tested the proposed approach employing a number of human models wearing various clothes along with their different body shapes, and the significance of the method to human posture recognition has been demonstrated.