Anomaly Detection in Cooperative Vehicle Perception Systems Under Imperfect Communication

Ashish Bastola, Hao Wang, Abolfazl Razi

Abstract: Anomaly detection in autonomous driving is crucial for safety and traffic efficiency. Traditional single-vehicle systems face limitations due to restricted visibility and occlusions in complex scenarios. Cooperative Perception improves accuracy by sharing sensor data between vehicles, but challenges such as bandwidth constraints and unstable connections can lead to delayed or missing information. In this work, we present Cooperative Perception-based Anomaly Detection (CPAD), a robust framework that can operate under imperfect communication. CPAD utilizes a graph-transformer architecture to model spatiotemporal correlation among vehicles and achieves a superior F1 gain of 15% and an AUC gain of 5% compared to conventional models. We also observe significant robustness against communication interruptions with an F1 gain of 19% under 25% sequential blackouts and 13% gain under 25% random blackouts compared to the best-performing model. Additionally, we present a benchmark dataset comprising 15,000 multi-agent scenarios with 90,000 vehicle trajectories. This dataset aims to address the current gap in multi-agent anomaly detection research and enhance the safety and reliability of autonomous driving systems. Code: https://github.com/abastola0/CPAD

External IDs:dblp:conf/urai/BastolaWR25