Quantification and classification of locomotion patterns by spatio-temporal morphable models

Morphable models have been applied successfully in the context of computer vision and computer graphics for the representation of classes of stationary images. We develop a similar technique for the representation of classes of complex movements that we call space-time morphable models. This technique permits one to approximate new complex movement patterns by linear combinations of few learned prototypical example patterns. The weights of the linear-combination provide a low-dimensional description of the patterns that can be exploited for the classification of the underlying actions, and also for the estimation of continuous parameters that quantify characteristic properties of the movement. (Examples are the direction of locomotion and the style with which a certain movement is executed.) We demonstrate the applicability of the technique for the classification and quantification of properties of locomotion patterns. Several possible applications of space-time morphable models in the context computer vision and surveillance are discussed.