EmbTCN-Transformer: An Embedding Temporal Convolutional Network–Transformer Model for Multi-Trajectory Prediction

This paper addresses the multi-trajectory prediction problem and a so-called Embedded-TCN-Transformer (EmbTCN-Transformer) model is designed by using the real-time historical trajectories in a formation. A temporal convolutional network (TCN) is utilized as the input embedding to introduce temporal awareness capabilities into the model. Then, the self-attention mechanism is incorporated as the backbone to extract correlations among different positions of the trajectory. An encoder–decoder structure is adopted to generate future trajectories. Ablation experiments validate the effectiveness of the EmbTCN-Transformer, showing that the TCN-based input embedding and the self-attention mechanism contribute to 30% and 80% reductions in prediction error, respectively. Comparative experiments further demonstrate the superiority of the proposed model, achieving at least 60% and 10% performance improvements over Recurrent Neural Network (RNN)-based networks and the conventional Transformer, respectively.