Deep Reinforcement Learning-Powered Extended Vehicular Cooperative Perception
Résumé
Cooperative perception (CP) in vehicular networks allows vehicles to share data and extend awareness beyond onboard sensors, a key requirement for advanced driving technologies to ensure safe and efficient driving decisions. In this paper, we propose a deep reinforcement learning (DRL)-based mechanism for selectively relaying Cooperative Awareness Messages (CAMs) through RoadSide Units (RSUs). Unlike prior work that focuses on local V2V exchanges or naïve RSU broadcasting, we propose the first deep contextual multi-armed bandit (MAB) framework for relevance-aware relaying at the RSU level. Our approach leverages vehicle kinematics, road topology, and spatio-temporal features to assess the importance of each message.
The objective is to enhance perception while minimizing bandwidth usage. Simulation results in a realistic urban scenario show that our method achieves up to 88% reduction in bandwidth compared to naïve broadcasting and 48% compared to a linear classifier, while achieving an accuracy of 96%. This framework is compliant with ETSI ITS-G5 Standard and provides a scalable foundation for intelligent cooperative perception.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| licence |
