In recent years, data analysis in sports has become increasingly popular, and recognizing in-game actions plays a crucial role in tactical planning. For example, obtaining a shot chart that visualizes where shots were taken and their success rates can help determine optimal defensive positioning to block opponents more effectively. However, continuous data collection from daily practices and games is essential for analysis, and the manual labeling process poses a significant challenge. Therefore, to automate labeling, this study aims to recognize in-game actions.
Decision-making in player actions is influenced by factors such as player position, orientation, and movement speed, which are all centered around court dynamics. Thus, we design an action recognition model that incorporates positional information.
For instance, in basketball, when a player possesses the ball, their actions vary significantly depending on whether they are in their own half, near the half-court line, or under the basket. The farther they are from the basket, the more likely they are to prioritize passing or dribbling over shooting. Given video data as input, our model first estimates player positions and extracts positional features.
Next, in parallel with position estimation, an existing action recognition model is employed to recognize individual movements. Since these movements are centered on human motion, we consider this as a gesture recognition task. Finally, the outputs from both processes are combined and passed through a classifier to determine the final action category.
Published Papers
- Ryosuke Hasegawa, Akira Uchiyama, Fumio Okura, Daigo Muramatsu, Issei Ogasawara, Hiromi Takahata, Ken Nakata, and Teruo Higashino, "Close-Contact Detection Using a Single Camera for Sports Considering Occlusion", IEEE Access, Vol. 10, pp.15457-15468, 2022
