CCB-GraphGAN: A Graph GAN-Based Approach for Lane-Level Traffic Anomaly Detection on Freeways

Abstract: Detecting anomalous events on freeways is challenging due to the fast-paced, high-volume nature of freeway traffic and the unpredictability of incidents such as crashes and breakdowns. Early detection of these anomalies is crucial for improving emergency response times and mitigating congestion. However, current traffic monitoring systems struggle to detect lane-level anomalies in a timely way, which hinders effective traffic management.
This study frames freeway anomaly detection as identifying anomalous nodes within graphs that represent spatiotemporal traffic variations across lanes. It introduces a Cycle-Consistent Bidirectional Graph Generative Adversarial Network (CCB-GraphGAN) to enhance the accuracy and timeliness of lane-level anomaly detection. The CCB-GraphGAN builds on a Bidirectional Generative Adversarial Network (BiGAN) and utilizes Graph Attention layers to model relationships between traffic flows in both observed and latent spaces. The model learns the joint distribution of normal traffic and its latent representations through adversarial training and incorporates two cycle consistency constraints to reliably reconstruct both normal traffic data and its latent features. A trained CCB-GraphGAN can accurately reconstruct regular traffic states; however, it exhibits high reconstruction errors in the presence of anomalies. During inference, these reconstruction errors are analyzed to identify anomalous nodes related to traffic incidents.
The proposed method is evaluated using the Freeway Traffic Anomalous Event Detection (FT-AED) dataset. The results demonstrate its superiority over existing baseline models in terms of both detection accuracy and speed. Notably, the CCB-GraphGAN model achieves a lower percentage of missed crashes compared to competing methods and significantly reduces detection delay. It identifies crashes an average of 5 minutes earlier than official reports while maintaining a remarkably low false positive rate of 1%. Additionally, an analysis of real-world crash events further validates the model’s ability to detect lane-level freeway anomalies early.