B5GCyberTestV2X Cooperative Perception: A Co-Simulation Framework Bridging Perception and Communication in Autonomous Driving

Cooperative perception (CP) in Vehicle-to-Everything (V2X) systems enhances autonomous driving awareness through multi-vehicle sensor data sharing. While state-of-the-art CP algorithms achieve superior performance under ideal communication assumptions, their viability under realistic 5G network constraints remains largely unexplored. This paper presents a rigorous comparative analysis of CP data sharing strategies-intermediate fusion with a state-ofthe-art attention-based model versus late fusion with fused object lists-under realistic V2X network conditions. A novel bidirectional co-simulation framework, B5GCyberTestV2X, is utilized, which tightly integrates the CARLA sensor simulator and the OMNeT++/SIMU5G network simulator, enabling a closed-loop evaluation where network performance directly impacts vehicle safety outcomes. Through extensive Monte Carlo simulations of safety-critical scenarios, it is demonstrated that communication constraints, not perception algorithms, are the dominant bottleneck. While intermediate fusion achieves a 14.1% higher detection accuracy (AP@0.5) under low network load, it fails to meet the 100 ms latency requirement at medium densities (10 vehicles), resulting in a 3.2× increase in near-miss events compared to the lightweight late fusion strategy. ... mehrThe results reveal that late fusion remains the only viable strategy for deploying safety-critical CP applications today, maintaining sub-20 ms latency across all densities while achieving 14% lower AP than intermediate fusion under ideal conditions but 3.2 × fewer near-miss events under realistic network load, and shows, through closed-loop simulation, that communication-induced latency dominates safety outcomes in urban V2X scenarios.