Possibility of a simplified support system for judging valid strikes in kendo using deep learning

Video judging has been introduced in a variety of sports, including VAR in soccer and challenge systems in volleyball. In the field of martial arts, video judging was introduced on a full scale in judo in 2007. In kendo, three cameras were used to make video judgments, and judgments were made by three to five judges, including the chief referee. However, it has yet to be put into practical use, having only been introduced in three tournaments. Existing studies have attempted to make highly accurate judgments based on auditory and visual information by using sensors and microphones or by shooting from directly above. However, the locations where these devices can be used are limited, making it difficult to introduce them to small-scale local tournaments and university sports. Therefore, in this research, we conduct learning and evaluation of kendo using ResNet-18 with camera images taken from the actual referee’s position as input. Then, we then verify the feasibility of the proposed method by obtaining effective hits using camera images from multiple directions.