Recognition of sumo attack styles based on the features between sumo wrestlers

Sumo is a high-contact combat sport contested between two sumo wrestlers, who wear a belt-like garment called a mawashi and grapple with each other. One of the key elements of analysis in sumo is the attack styles, known in Japanese as torikuchi, which refer to the tactical body movements and engagement strategies used to gain an advantage. Attack styles reflect both the overall strategy of the bout and the individual characteristics of the sumo wrestlers, and they change depending on the situation within the match. There is no standard official record of these changes, which makes quantitative analysis of the bout strategy difficult. To address this challenge, we propose a method for automatically recognizing attack styles from sumo videos. From the sumo videos, we estimate the wrestlers’ joint points at the moments when the attack styles are executed using OpenPose. The features based on joint points are used as inputs, and the attack style names are used as annotations. The model utilizes an LSTM and a Transformer architecture. We pick up two popular attack styles: oshizumo and yotsu-zumo.