Research map / Detection and segmentation / Open-set and few-shot detection
FT-AED: Benchmark Dataset for Early Freeway Traffic Anomalous Event Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Releases FT-AED, a lane-level radar traffic dataset from I-24 near Nashville with crash logs and expert anomaly labels, and benchmarks graph and sequence autoencoders for early crash detection before official reporting.
Problem
Freeway anomaly (crash) detection suffers from delayed and error-prone manual incident reports, and existing large freeway traffic datasets lack anomaly labels and lane-level, fine-grained temporal resolution.
Approach
Provide a lane-level radar sensor dataset sampled every 30 seconds with crash labels and expert-labeled anomalies, and benchmark reconstruction-based autoencoders (GCN, GAT, RGCN spatiotemporal graphs, GCN-LSTM, Transformer, MLP) trained on anomaly-free data, measured by a Reduction in Reporting Delay metric.
Key results
- GCN autoencoder achieved the lowest mean reporting delay (-10.20 ± 5.98 minutes) and highest AUC (0.70) among the compared models at 5% FPR, detecting crashes over 10 minutes before official reports on average.
- STG-GAT AE at 10% FPR missed 0 of the 44 reported crashes with a mean delay of -13.33 ± 2.56 minutes on the full dataset.
- Removing manually labeled non-crash anomalies from training reduced reporting delay from -8.45 to -10.20 minutes at 5% FPR, though miss percentage rose from 8.3% to 25%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use spatial graph relationships instead of treating each node independently | GCN AE versus Transformer AE and MLP AE · FT-AED validation | AUC: 0.60/0.62 -> 0.70 | Helped |
| move temporal information into the spatiotemporal graph rather than process it with an LSTM | Spatiotemporal graph autoencoders versus GCN-LSTM AE · FT-AED validation | AUC: 0.65 -> 0.67/0.68 | Helped |
| use spatiotemporal GAT rather than GCN at 10% FPR | STG-GAT AE versus GCN AE · FT-AED validation (10% FPR) | Miss Percentage: 16.67% -> 0% | Helped |
| include manually labeled anomalies in training data (ablation) | GCN AE · FT-AED validation (5% FPR) | Miss Percentage: 25% -> 8.3% | Helped |
| include manually labeled anomalies in training data (ablation) | GCN AE · FT-AED validation (5% FPR) | Reporting Delay: -10.20±5.98 -> -8.45±7.81 | Hurt |
| include manually labeled anomalies in training data (ablation) | GCN AE · FT-AED validation (5% FPR) | AUC: 0.70 -> 0.69 | Hurt |
Surprising findings
- Removing manually labeled anomalies from training lowered detection performance on miss percentage, suggesting that some anomalies in training improve robustness.
- The Transformer autoencoder performed worst, with positive mean reporting delay (+2.42 min), indicating it detected crashes after they were reported.
Limitations
- Validation relies on only 44 officially reported crashes and uncertain crash timing, so results are noisy with high variance.
- Validation is only at the crash-report level; anomalies are not validated at the lane or location level.
- Only reconstruction-based autoencoder methods are benchmarked; AUC may be misleading for instantaneous anomaly labels.
Experimental setup
Datasets: FT-AED
Benchmarks: FT-AED
Metrics: Reduction in Reporting Delay Miss Percentage Reconstruction MSE False Positive Rate AUC
Baselines: GCN AE GCN-LSTM AE Transformer AE MLP AE
Builds on: Relational Graph Convolutional Network Graph Attention Network Graph Convolutional Network Tree-Structured Parzen Estimator Adaptive Smoothing Method
Closest papers by meaning
- TraffiDent: A Dataset for Understanding the Interplay Between Traffic Dynamics and Incidents (NeurIPS 2025 · similarity 0.86)
- When Every Millisecond Counts: Real-Time Anomaly Detection via the Multimodal Asynchronous Hybrid Network (ICML 2025 · similarity 0.85)
- EchoTraffic: Enhancing Traffic Anomaly Understanding with Audio-Visual Insights (CVPR 2025 · similarity 0.83)
- Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving (CVPR 2025 · similarity 0.82)
- Self-Supervised Likelihood Estimation with Energy Guidance for Anomaly Segmentation in Urban Scenes (AAAI 2024 · similarity 0.82)
- eTraM: Event-based Traffic Monitoring Dataset (CVPR 2024 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.