Learning and comparing trajectories with a GNG-based architecture

Abstract

In this paper we use a kind of self-organising network, the Growing Neural Gas, as structure capable of characterizing hand posture, as well as its movement. Topology of a self-organizing neural network determines posture, whereas its adaptation dynamics throughout time determines gesture. This adaptive character of the network allows us to avoid the correspondence problem of other methods, so that the gestures are modelled by the movement of the neurons. Using this representation to an image sequence we are able to follow the evolution of the object along the sequence learning the trajectory that describes. Finally we use the Hausdorff distance among trajectories to compare and recognize them. © 2009 Springer-Verlag Berlin Heidelberg.