← All work

Lane-Level Freeway Traffic Anomaly Detection

A graph-based generative adversarial model for early detection and localization of anomalous traffic events across freeway lanes, while minimizing false alarms.

Anomaly detection · Graph neural networks · Generative models
Speed heatmap along Interstate 24 showing the detected traffic anomalies before the official crash report, alongside the corresponding reconstruction error.
A crash on Interstate 24: (a) lane speeds with the model's anomaly alerts, showing detection before the official report; (b) the model's reconstruction error. From Nouri & Ilbeigi, IEEE Access 2026 (CC BY 4.0).

The problem

Detecting freeway incidents at an early stage remains challenging, particularly when disruptions initially affect only a small portion of the roadway. Conventional monitoring systems often rely on surveillance cameras, manual observation, and incident reports, which can introduce delays in identifying emerging disruptions.

High-resolution radar networks provide detailed traffic measurements across individual lanes, creating opportunities for more precise incident detection. However, identifying abnormal conditions requires models that can capture spatial dependencies between neighboring lanes and road segments while distinguishing unexpected disruptions from normal traffic variations.

Data

Interstate 24 · Nashville, Tennessee

49radar sensors along the corridor
18 miwestbound freeway segment
196graph nodes: 4 lanes × 49 mile markers
30 ssampling interval

The measurements come from the FT-AED (Freeway Traffic Anomalous Event Detection) dataset: speed, occupancy, and traffic volume for each lane, recorded in October 2023, with labeled crash events. The labeled crashes were used only for evaluation.

Approach

Graph-based traffic representation

Freeway traffic is represented as a graph in which each node corresponds to a specific lane at a given mile marker. Nodes are connected to neighboring lanes at the same location and to lanes at adjacent mile markers.

This structure allows the model to capture spatial dependencies across the freeway network rather than treating individual sensor measurements independently.

Freeway graph visualization showing lane-level nodes across mile markers, with connections between adjacent lanes and longitudinally neighboring road segments.
The freeway graph: one node per lane at each mile marker, linked to adjacent lanes and to neighboring road segments. Node color shows speed. From Nouri & Ilbeigi, IEEE Access 2026 (CC BY 4.0).

CCB-GraphGAN architecture

CCB-GraphGAN (Cycle-Consistent Bidirectional Graph Generative Adversarial Network) is proposed as an extension of the Bidirectional GAN framework for anomaly detection in graph-structured traffic data.

The architecture consists of three components, all implemented using Graph Attention Network (GAT) layers:

  • Generator: maps latent representations into synthetic traffic feature matrices, learning to reproduce the distribution of normal traffic conditions.
  • Encoder: maps observed traffic data into a compact latent representation, learning the inverse mapping of the generator.
  • Discriminator: distinguishes between real traffic–latent representation pairs and generated pairs through adversarial training.

Cycle-consistency learning

A key component of the proposed architecture is the introduction of two auxiliary cycle-consistency losses in the data and latent spaces. These constraints encourage the generator and encoder to learn consistent inverse mappings:

  • Data-space consistency: reconstructs the original traffic features after encoding and generation.
  • Latent-space consistency: recovers the original latent representation after generation and re-encoding.

Together, these constraints improve reconstruction fidelity and provide a more reliable basis for identifying deviations from normal traffic behavior.

CCB-GraphGAN architecture diagram showing the generator, encoder, discriminator, adversarial learning process, and cycle-consistency constraints.
CCB-GraphGAN: generator, encoder, and discriminator built from graph attention layers, with cycle consistency in both data and latent space. From Nouri & Ilbeigi, IEEE Access 2026 (CC BY 4.0).

Lane-level anomaly detection

During training, the model learns exclusively from normal traffic observations. During inference, the encoder and generator reconstruct each observed traffic state.

Anomalies are identified through reconstruction errors computed independently for each graph node. When a node’s reconstruction error exceeds its location-specific detection threshold, the corresponding lane is flagged as anomalous.

This enables the model to identify not only when abnormal traffic conditions emerge, but also where they occur across the freeway network.

Results

~5 minaverage detection lead over official crash reports, at a 1% false positive rate
30%+lower reconstruction error on normal traffic than the best baseline
9 msaverage inference time per graph snapshot on a CPU

Comparative evaluation

Compared with all evaluated baseline models, CCB-GraphGAN:

  • detected anomalies earlier;
  • missed fewer crashes at both 1% and 5% false positive rates;
  • achieved a higher area under the ROC curve (AUC);
  • reconstructed normal traffic more accurately.

Methods & tools

Graph attention networks · bidirectional GANs · cycle-consistency learning · anomaly detection · Python · PyTorch

This research was supported by the U.S. National Science Foundation under Grant No. CMMI-2026795.