Graph Multi-Resolution Transformer for Road Traffic Anomaly Detection

Abstract

Traffic anomaly detection has become increasingly vital for ensuring the safety, efficiency, and resilience of urban transportation systems. The emergence of megacities and the integration of autonomous vehicles (AVs) introduce new challenges for anomaly detection in mixed-traffic environments. AVs exhibit distinct driving behaviors compared to human drivers, resulting in more complex traffic scenarios that necessitate advanced detection methods. This paper presents a novel Graph Multi-Resolution Transformer (GMRT) model designed for traffic anomaly detection in environments where autonomous and human-driven vehicles coexist. Utilizing simulation data from the autonomous driving demonstration zone in South Korea, the GMRT effectively captures temporal patterns at multiple resolutions and leverages graph-based spatial relationships to enhance anomaly detection performance. Unlike traditional models that focus on macro-level traffic flow, our approach operates at a micro-level (lane-based), enabling precise identification of anomalies on a per-lane basis. Comparative experiments against established time-series models demonstrate the superior accuracy and robustness of GMRT, particularly in early anomaly detection across varied temporal horizons. The results underscore the model’s potential as a robust solution for contemporary and future urban traffic management challenges, especially in scenarios involving the increasing penetration of AVs.

Keywords

• Deep learning
• intelligent transportation monitoring
• road traffic anomaly detection
• multi-resolution learning
• time-series analysis