Generic temporal segmentation of cyclic human motion

A method is proposed for the temporal segmentation of cyclic human motion from video sequences. The proposed method is divided into three processing steps. Once silhouettes and body part locations are obtained, a set of individual 1-D signals representing motion trajectories of body parts is extracted for the entire sequence. The second step performs the individual segmentation of all signals in the set in order to localize their periodic segments. In the final step, all individual segmentations are coherently merged into a global segmentation for the entire sequence and set of signals. The proposed approach has been successfully tested on a variety of sequences containing cyclic activities such as aerobic exercises and walking along different directions.