Parametrized SOMs for hand posture reconstruction

Abstract

This paper describes the use of neural network for gesture recognition based on finger tips, a system that recognizes continuous hand postures from video images. Our approach yields a full identification of all finger joint angles. This allows a full reconstruction of the 3D hand shape, using an artificial hand model with 16 segments and 20 joint angles. The focus of the present paper is how to employ a parametrised SOM neural network for the inverse kinematics task to compute the angles of a hand model out of 3D positions of the fingertips. We show that this type of neural net does not only achieve excellent results from very few training examples, but also can be applied to uncommon data structures.