Spatio-Temporal Context Graph Transformer Design for Map-Free Multi-Agent Trajectory Prediction

Predicting the motion of surrounding vehicles is an important function of autonomous vehicles. However, most of the current state-of-the-art trajectory prediction models rely heavily on map information. In order to overcome the shortcomings of the existing models, our paper proposes a map-free trajectory prediction model and names it TR-Pred (Trajectory Relative two-stream Prediction). The trajectory stream employs LSTM to embedding the trajectory information of each agent. Subsequently, it utilizes graph neural networks (GNN) to extract latent traffic information in the current scenario, such as lane lines, drivable areas, and traffic control conditions. The relative stream utilizes temporal transformer to capture the local relative movement among agents. Subsequently, it employs GNN to extract the interaction information of all target agent. We augment the temporal transformer through refine initialization of its class token. This refined thereby enable enhanced modeling of inter-agent relative motion correlations between the agents. The decoder predicts the target agent by incorporating the historical interaction information among agents with latent traffic information. We validate TR-Pred on the Argoverse dataset, the highD dataset and the rounD dataset. The results show that TR-Pred performs better in the minimum Average Displacement Error compared to the main map-base model use in 2020, 2021. Experiments on Argoverse results show that our framework achieves a 16.3%/19.7%/24.4% improvement in minADE/minFDE(minimum Final Displacement Error)/MR(Miss Rate) compared to CRAT-Pred. The experimental results on highD and rounD show that, compared to the map-free version HiVT, our framework achieves improvements of 18.8%/18.8%/25.0% and 8.3%/8.5%/7.4% in minADE/minFDE/MR, respectively.