Person identification from spatio-temporal volumes

This paper proposes a novel identification technique for a person from gait and body shape. Although the shape of onepsilas body has not been considered much as a characteristic for the person identification, it is closely related to gait and it is difficult to disassociate them. The proposed technique utilizes the full spatio-temporal volume carved by a person who walks and the average image created from the spatio-temporal volume. Affine moment invariants are derived from the spatio-temporal volume and the average image, and classified by a support vector machine. Experiments using a standard gait database show this method may produce better results than those based on gait analysis alone and k-nearest neighbor classification.