Fully Interactive Graph-Based Trajectory Prediction via Topological Scenario Representation

The accurate trajectory prediction for vehicles and other traffic agents is essential for the safety and efficiency of transportation environment construction. However, the trajectory prediction task can be affected by many factors such as road constraints, vehicle intentions, interactions with nearby agents and so forth, which makes the prediction challenging and time-consuming. To address the complex traffic conditions and heterogeneous impact factors, this study proposes a fully interactive graph-based trajectory prediction method with the topological scenario representation. Specifically, the traffic scenario is firstly constructed as a topological graph to maintain the spatial relationship among agents and map. The temporal features of traffic states are then obtained via a Gated Recurrent Unit processor. After that, two types of interaction graph are generated based on the topological scenario and a directed edge-enhanced graph network is adopted for the extraction of both inter-agent and agent-map interactive features. Finally, a Graph Convolutional Network block is employed to encode the whole scenario context information. A Long Short-Term Memory decoder is used for future trajectory generation based on the above spatial-temporal interactive features. The proposed model is trained and validated on Argoverse2 dataset, and the results demonstrate the effectiveness of our approach.