Trajectory prediction for heterogeneous road-agents using dual attention model

We propose a Dual Attention model (DATP) for predicting the near-term trajectories of road agents. Our model includes an interaction attention module and a trajectory attention module. In the interaction attention module, a new hybrid CNN-LSTM network model is designed to model the interaction between different road agents and the attention mechanism is introduced to implicitly quantify the dynamic influence of interaction on the future trajectory of road-agents. The trajectory attention module uses an improved-LSTM network to encode the historical trajectory information. Then, the trajectory information and interaction information are fused into the prediction network to obtain the predicted near-term trajectory. We evaluate the performance of our model on the standard datasets. Comparison results indicate that DATP could achieve a more accurate prediction trajectory over 35% for heterogeneous road-agents. Furthermore, the results of ablation experiments show that our model has great contributions to the improvement of trajectory prediction accuracy.