Fig. 5
GNN-GRU model architecture. (A) The implementation principle of the GNN, where ni represents nodes (joints), ej, sj, rj represent the set of edge features (bones), \(e_{j}\) represents the edge features, and sj, rj represent the start and end nodes of an edge, respectively. g denotes the global features of the graph. \(f_{g}\), fn, fe represent the global update function, node update function, and edge update function, respectively. GN1, GN2, and GN3 represent the three convolutional layers, GRU represents the gated recurrent unit layer, and Linear represents the fully connected layer. (B) The hybrid model used in this study, where Node features represent the feature matrix, Target variables represent the pitching speed matrix, and Edge index represents the edge index matrix. (C) Architecture of gated recurrent unit.
