This work deals with the problem of representing and recognizing human body movements, given XYZ tracking data. Prior approaches by other researchers used a smaller number of classification categories, which demanded less attention to representation.
We develop techniques for representation of movements based on space curves in subspaces of a ``phase space.'' The phase space has axes of joint angles and torso location and attitude, and the subspaces are subsets of the axes of the phase space. Using this representation we develop a system for learning new movements from ground truth data by searching for subspaces in which the movement to be learned describes a curve which is easily separated from other movements. We then use the learned representation for recognizing movements in data.
We train and test the system on nine fundamental movements from classical ballet by two dancers, and show the system can learn, recognize and segment out five of the movements accurately, but confuses one pair of movements from one dancer and another pair from the other dancer. Finally, we suggest how the system can be improved and extended.