Recognizing human motion using eigensequences

Abstract

This paper presents a novel method for motion recognition. The approach is based on 3D motion data. The captured motion is divided into sequences, which are sets of contiguous postures over time. Each sequence is then classified into one of the recognizable action classes by means of a PCA based method. The proposed approach is able to perform automatic recognition of movements containing more than one class of action. The advantages of this technique are that it can be easily extended to recognize many action classes and, most of all, that the recognition process is real-time. In order to fully understand the capabilities of the proposed method, the approach has been implemented and tested in a virtual environment. Several experimental results are also provided and discussed.