Explainable multimodal trajectory prediction using attention models

Automated vehicles are expected to navigate complex urban environments safely along with several non-cooperating agents. Therefore, accurate trajectory prediction is crucial for safe nav-igation and receives much interest. However, most existing studies mainly focus on models' prediction accuracy rather than their explainability. In this paper, we propose a Multimodal Trajectory Prediction Transformer (MTPT) model to retrieve the influencing factors of prediction and help understand the intrinsic mechanism of prediction. Especially in the MTPT model, we use a modified Swin Transformer with multiple prediction heads to carry out multimodal trajectory prediction. Numerical experiments confirm the MTPT model can capture the most critical input factors with the help of the attention method and thus improve prediction accuracy. The proposed model obtains state-of-the-art results even with limited training data. Moreover, the identified input factors are also in agreement with the human driving experience. This agreement indicates that the proposed model appropriately learns how to predict.