PhysicalAI Traffic Anomaly Reasoning: AI City Challenge 2026 Training Data

44,040 pseudo-labeled multi-task annotations cover 3,670 CCTV transportation videos, totaling approximately 26.1 hours of footage. This dataset is the official training split for the AI City Challenge 2026 Track 3 — Anomalous Events in Transportation. It was created by NVIDIA and last updated in May 2026.

Use Cases

• Training anomaly detection models based on CCTV video footage.
• Developing multi-task reasoning systems based on the 10 described annotation types.
• Benchmarking video understanding models for the AI City Challenge.
• Analyzing temporal patterns in anomalous traffic events.

Strengths

• 44,040 annotations provide a substantial training corpus.
• 3,670 videos offer approximately 26.1 hours of real-world footage.
• Annotations span 10 distinct task types for multi-task learning.
• Sourced from eight public datasets, potentially increasing diversity.

Limitations

• Annotations are pseudo-labeled, which may introduce label noise.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset's specific geographic and temporal coverage is not detailed.

Provenance

Source

NVIDIA, for the AI City Challenge 2026.

Collection Method

Pseudo-labeled annotations sourced from eight public datasets.

Time Range

The temporal coverage of the source videos is not specified.

Freshness

Last updated 2026-05-03 15:05:23.

Geography

The geographic origin of the CCTV videos is not specified.