MoCVAE: Movement Prediction by A Conditional Variational Autoencoder for Doubles Badminton

AI has emerged as a potent tool of sports analysis, but there is no research yet for doubles badminton. In this paper, we introduce a novel movements prediction model, namely MoCVAE, for doubles badminton. MoCVAE differentiates the doubles setting from the singles and leverages its uniqueness to improve performance. Specifically, the Player Position Embedding module is pre-trained with singles data to initialize embeddings to overcome the scarcity of data. On top of a CVAE-based structure, a novel Motion Interaction module further employs a GAT to refine the embeddings based on interaction among the players. Finally, in addition to player movements, MoCVAE simultaneously predicts hitting players and shot types since they are all significantly correlated. The experimental results on real datasets show that MoCVAE outperforms its variants by 61.22% in terms of ADE, manifesting that MoCVAE integrates all the essential perspectives successfully.