Rapidly identifying anomalous events on freeways can significantly improve emergency response time and reduce congestion. However, the unpredictable nature of these incidents, along with the limitations of current monitoring systems, makes accurate and timely detection challenging. In this study, we propose a Cycle-Consistent Bidirectional Graph Generative Adversarial Network (CCB-GraphGAN) model to enhance lanelevel traffic anomaly detection on freeways. By treating freeway traffic information as graph-structured data, our model captures the joint distribution of regular traffic patterns and their underlying latent representations through an adversarial process. It also incorporates cycle consistency constraints to effectively reconstruct both normal traffic data and its latent features during training. During the inference phase, we employ an autoencoder-based anomaly detection approach to identify lanelevel traffic anomalies. We evaluate the performance of our method through comparative and case study analyses using the recently released Freeway Traffic Anomalous Event Detection (FT-AED) dataset. The results indicate that our method outperforms all baseline models in detecting freeway anomalous events early. Notably, it achieves a lower percentage of missed crashes compared to baseline models. With a false positive rate (FPR) of only 1%, our method reduces detection delay by identifying crashes an average of 5 minutes earlier than their official report time. Additionally, an analysis of real-world crash events further confirms our method's effectiveness in accurately identifying freeway anomalies at an early stage.

Abnormal traffic patterns caused by extreme events have the potential to disrupt traffic flow on large regions of urban road networks. Timely and reliable detection of such disruptions is crucial for effective traffic management. However, existing methods for detecting extreme traffic events are unable to simultaneously identify disruptions at both network and local levels. Moreover, most existing methods cannot handle incomplete traffic data, which can be an essential challenge during extreme events. In this study, we address these limitations by creating a novel method for traffic anomaly detection using a multitask sequence-to-sequence learning process. Specifically, we propose a Self-Imputing Deep Multitask Sequence (SI-DMSeq) model that can simultaneously reconstruct sequential traffic data and provide one-step-ahead predictions. This model also integrates an unsupervised data imputation strategy to enable training with incomplete data. During inference, the proposed model adopts an autoencoder-based anomaly detection approach to identify large-scale disruptions at both network and local levels. Additionally, it utilizes one-step-ahead predictions to reliably impute missing data under typical traffic conditions. To evaluate the performance of our proposed method, we applied it to the Manhattan road network in New York City using historical traffic data. The results indicated that the proposed method accurately detects disruptions caused by Hurricane Sandy in 2012 at both network and local levels and efficiently imputes missing traffic data under typical conditions.