Affiliation:
1. Wuhan University of Technology
2. University of Waterloo
Abstract
<div class="section abstract"><div class="htmlview paragraph">Accurately predicting the future trajectories of surrounding traffic agents is
important for ensuring the safety of autonomous vehicles. To address the
scenario of frequent interactions among traffic agents in the highway merging
area, this paper proposes a trajectory prediction method based on interactive
graph attention mechanism. Our approach integrates an interactive graph model to
capture the complex interactions among traffic agents as well as the
interactions between these agents and the contextual map of the highway merging
area. By leveraging this interactive graph model, we establish an agent-agent
interactive graph and an agent-map interactive graph. Moreover, we employ Graph
Attention Network (GAT) to extract spatial interactions among trajectories,
enhancing our predictions. To capture temporal dependencies within trajectories,
we employ a Transformer-based multi-head self-attention mechanism. Additionally,
GAT are utilized to model the interactions between traffic agents and the map.
The method we propose comprehensively incorporates the influences of time,
space, and the map on trajectories. The interactive graph models can serve as
effective prior knowledge for learning-based approaches, thereby enhancing the
acquisition of interaction patterns among traffic scenarios and facilitating the
interpretability of the method. We evaluate the performances of our method using
real-world trajectory datasets from the highway merging area, i.e., the Exits
and Entries Drone Dataset (<i>exiD</i>). Comparative analysis against
classical algorithms demonstrates a reduced trajectory prediction error for
prediction horizons of both 3s and 4s.</div></div>