Abstract
In this paper, a method of increasing the realism of non-verbal interaction in 3D is proposed. The method is aimed at creating believable truly autonomous actors in virtual environments. A simple action-reaction interaction model with a predefined set of actions is used. The actions are recognised using weighted-directed graphs (Markov Models), using a set of control points as features to reduce the dimensionality of the recognition task, thus enabling fast and effective recognition. The method naturally implements characteristics such as reaction delay and gradual recognition with possible mistakes. The approach is tested on a simplified boxing scenario.