VRU-Accident: Video Benchmark for Vulnerable Road User Safety

VRU-Accident is a video dataset for benchmarking AI on traffic safety, created by kyh9191 and last updated on July 8, 2025. It contains 6,000 samples for Video Question Answering (VQA) across six categories and 1,000 samples for dense captioning. The VQA task includes one correct answer and three contextually relevant counterfactuals per question.

Use Cases

• Train Video Question Answering models based on the six defined categories (Weather & Light, Traffic Environment, etc.).
• Benchmark dense video captioning models on accident scenarios.
• Evaluate model robustness using the provided counterfactual answer options in the VQA task.
• Study causal reasoning in traffic accidents based on the structured question categories.

Strengths

• Provides 6,000 VQA samples with 24,000 candidate answer options.
• Includes a distinct task for dense captioning with 1,000 samples.
• VQA questions are structured into six specific, safety-relevant categories.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the overall dataset is unknown, which may limit suitability assessment.

Provenance

Source

kyh9191 on Hugging Face.

Freshness

Last updated 2025-07-08 17:44:37; freshness should be verified.