Improved Informer for Pose Prediction of Badminton Sports

For badminton sports, the close proximity of the initial sequences of certain movements makes it difficult to predict the subsequent movements of badminton players. In order to solve this problem, we propose a variant of the Transformer architecture based on the self-attention mechanism, which adds a classification module based on Informer model. We also introduce a hyperparameter item to balance prediction loss and classification loss. We use Kinect to collect badminton motion data to build our own dataset for model training. A series of comprehensive experiments have demonstrated that the enhanced Informer exhibits a 4.3% and 2.8% improvement in performance on the NTU and Badminton datasets, respectively.