Dynamic attack motion prediction for kendo agent

Abstract

A motion prediction method using Gaussian Mixture Models (GMM) is applied to a kendo agent (Kendo is a traditional Japanese martial art). Human player motion is measured by a motion capture system, using markers attached to each of the player's joints. Measurement information is converted to a state vector with Euler angles to indicate orientation of the sword and orientation of each part of the player's body. To model the motion as a nonlinear dynamical system, GMMs are generated from a demonstration set when an opponent is attacked. The efficiency of the proposed method is experimentally verified.