Roadside Unit-assisted Cooperative Perception at Hazardous Intersections: A Digital Twin-Based Safety Evaluation Framework

Abstract

Occlusion-induced perception limitations at intersections remain a major safety challenge for connected and automated vehicles (CAVs), particularly when buildings, roadside structures, or surrounding traffic block the vehicle’s onboard sensors. This paper presents a roadside unit (RSU)-assisted cooperative perception system and a digital twin-based safety evaluation framework for real-world hazardous intersections. The RSU detects and tracks vehicles from an elevated LiDAR viewpoint, converts the results into lightweight object-centric states, and delivers them to an Autoware-based vehicle stack via vehicle-to-infrastructure (V2I) communication without modifying the planner. We compare onboard-only and RSU-assisted conditions at three site-specific intersections, with 100 runs per site and condition. Safety impact is evaluated offline using time-to-collision (TTC)/deceleration rate to avoid a crash (DRAC)-based surrogate safety measures (SSMs) and conflict-relevant safety-critical pass rates. The results show that RSU assistance reduces verified safety-critical pass rates across all sites, with the largest improvement under dynamic occlusion, where the verified TTC-based risk decreases from 37% to 2% and the verified DRAC-based risk decreases from 20% to 2%. These findings suggest that RSU-assisted cooperative perception can mitigate occlusion-induced risks, while revealing that the benefit is determined by the occlusion mechanism at each site and that conflict-based filtering is required to separate genuine risk from routine yielding-a distinction perception-level evaluation cannot make.