Precise trajectory prediction in complex driving scenarios is crucial for autonomous vehicles. To achieve so, the autonomous vehicle has to effectively model the interactions between surrounding vehicles and consider the environmental information at the same time. This research proposes a unified prediction framework that simultaneously models the vehicle-vehicle interaction and the vehicle-environment relation. With a heuristic scene representation in graph structure in which vehicles and spaces are modeled as nodes, a dual graph attention module is adopted to model repulsion and attraction respectively from surrounding vehicles and the non-vehicle spaces. Deterministic trajectories are predicted using the updated repulsion- and attraction-aware scene encodings. Experiments are conducted on the NGSIM dataset to show the effectiveness of the proposed framework. Maneuver-based statistics are collected to analyze the prediction ability of the proposed model facing different maneuver trajectories.
RA-GAT: Repulsion and Attraction Graph Attention for Trajectory Prediction
19.09.2021
566459 byte
Aufsatz (Konferenz)
Elektronische Ressource
Englisch
Segmentation with pairwise attraction and repulsion
IEEE | 2001
|Segmentation with Pairwise Attraction and Repulsion
British Library Conference Proceedings | 2001
|Attraction in repulsion for maglev vehicles
Tema Archiv | 1975