One essential task for autonomous driving is to accurately predict the future motions of surrounding traffic agents. Recently, Graph Neural Network (GNN) approaches have shown potential in motion prediction due to the fact that information in the traffic scenario can be inherently formed into a graph structure. However, existing approaches are limited to using a static graph representing the states of the current timestamp, ignoring the prediction scenario’s derivation. In this work, we propose the Future Graph Network (FGNet) a two-stage GNN-based model for accurate and real-time multi-agent motion prediction. We design a future graph decoder that takes first-stage F mode multi-agent prediction result as input and reconstructs a graph to represent the derivation result of the scenario. The decoder enhances the first-stage result by (1) Constraining the agent’s future trajectory in the same modal to be harmonious and (2) Introducing the subsequent road information to guide the long-term prediction. Meanwhile, we incorporate statistical characteristics of the vehicles’ trajectory into the design of the loss function, which further boosts the performance of our model. Experiments show that FGNet achieves competitive performance on the Argoverse motion forecasting benchmark with real-time inference.
FGNet: A Graph-Based Motion Forecasting Method From a Future Perspective
04.06.2023
1832348 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
STGT: FORECASTING PEDESTRIAN MOTION USING SPATIO-TEMPORAL GRAPH TRANSFORMER
British Library Conference Proceedings | 2021
|