Application of AI to Kumite contests in Karate: Observations based on experiments to identify various attacking motions

Abstract

In Karate-Kumite matches, there are many cases where the referee’s judgment is difficult, and an erroneous judgment often occurs. In recent years, artificial intelligence,(maybe cut such as deep learning) has been introduced into ball games, athletics, competitive swimming, gymnastics, etc. It has been applied not only to judgments but also to tactics and training systems, but research reports on Karate-Kumite competition have not been reported yet. There are problems such as the inability to attach sensors to the body, and players cross each other at high speed, many blind spots occur. Therefore a method to get the motion data during competition and a Basic Method for identifying the motions has not been developed. Then, we propose a method to acquire motion data of Karate-Kumite and a basic method for identifying the movements. The method of recognizing play using (deep learning) AI in other sports such as soccer and ice hockey is based on achieving a certain degree of accuracy with a convolutional neural network (CNN).The output of CNN is connected to a system for more advanced play recognition. Therefore, in this study, we shot a Karate-Kumite imitating a match with a video camera and conducted a motion identification experiment using CNN. As a result, we obtained the required accuracy in order to connect to a system that performs more advanced motion identification with the estimation of “attack technique” or “not attack technique” or “attack technique hit” or “not hit.” In conclusion, from this motion identification experiment for Karate, the problem of introducing AI was clarified.