View-invariant hand-posture recognition for soft-remocon-system

The hand-posture-based user interface has been considered as an alternative to human-computer interaction (HCI), or for human-robot interaction, but many difficulties remain to be resolved. For natural recognition of hand posture by a vision device, we find that view-point-in variance is one of the most required characteristics. This paper presents a hand-posture recognition method that is robust against various viewpoints. Observing that when the viewing direction is changed, the same appearance feature does not necessarily match the 3D structure, and vice versa, we propose a 2-layered hand-posture database structure in which the 2D appearance-based method and the 3D-based method are complementarity combined. We have tested the approach in an experiment for 15 different viewing directions.