Abstract

In this paper, we introduce a tuned eigenspace technique so as to clas- sify human motion. The method presented here overcomes those problems related to articulated motion and dress texture effects by learning various human motions in terms of their sequential postures in an eigenspace. In order to cope with the variability inherent to articulated motion, we propose a method to tune the set of sequential eigenspaces. Once the learnt tuned eigenspaces are at hand, the recognition task then becomes a nearest-neighbor search over the eigenspaces. We show how our tuned eigenspace method can be used for purposes of real- world and synthetic pose recognition. We also discuss and overcome the problem related to clothing texture that occurs in real-world data, and propose a back- ground subtraction method to employ the method in out-door environment. We provide results on synthetic imagery for a number of human poses and illustrate the utility of the method for the purposes of human motion recognition.