HSTA: A Hierarchical Spatio-Temporal Attention Model for Trajectory Prediction

Predicting the future trajectories of surrounding agents has become an crucial problem to be solved for safety autonomous vehicles. Recent studies based on Long Short Term Memory (LSTM) networks have shown powerful abilities model social interactions. However, many these approaches focus spatial interactions neighborhood but ignore temporal that accompany In this paper, we propose a Hierarchical Spatio-Temporal Attention architecture (HSTA), which activates utilization with different weights, and jointly considers across time steps all agents. More specially, graph attention mechanism (GAT) is presented capture interactions, multi-head (MHA) conducted encode correlations state gated fusion (SGF) layer used integrate We evaluate our proposed method against baselines both pedestrian vehicle datasets. The results show effective achieves state-of-the-art achievements.