Achieving Level-4 autonomy in urban intersections through EKF-based multi-modal fusion enhanced by dual-attention PPO

초록

Autonomous vehicles require robust and context-aware decision-making to safely navigate complex urban intersections. To address challenges of perception uncertainty, communication delay, and multi-agent interaction, this paper proposes a novel framework combining multi-modal sensor fusion with confidence-weighted V2X message aggregation and dual-attention reinforcement learning. In the proposed system, RSUs employ an EKF to integrate LiDAR and camera data with CAM, CPM, and DENM messages over the 5G NR PC5 sidelink, generating a unified environmental representation with confidence weighting. This fused state is periodically broadcast to vehicles, where each onboard unit applies a dual-attention module to extract salient temporal and spatial features for policy learning. A Dual-Attention PPO (DA-PPO) agent then optimizes intersection maneuvers lane changing, collision avoidance, and traffic flow management using these context-rich inputs. Simulation results using the V2AIX dataset demonstrate that the proposed DA-PPO achieves up to 97.4% decision accuracy, 15%-20% higher packet-delivery reliability, and 2.3xfaster policy convergence compared with baseline A2C (PC5 interface) and PPO models. Furthermore, a decision-accuracy-based autonomy sublevel classification is introduced to benchmark high-autonomy decision performance with reference to SAE autonomy levels within the evaluated intersection scenarios. Overall, the proposed approach enables scalable, interpretable, and communication-aware autonomy for next-generation intelligent transportation systems.

키워드