VLM_traffic_incident

The dataset provides three complementary types of textual annotations: descriptive captions, question-answer pairs, and chain-of-thought reasoning, forming a progressive training structure from basic scene description to task-oriented supervision and causal inference.

The data are sourced from traffic surveillance cameras and dashcams, encompassing six event categories—two-wheeled vehicle intrusion, road-occupying construction, congestion, abnormal parking, spilled objects, and traffic accidents—across multiple countries with varied environmental and lighting conditions. This dataset serves as a high-quality resource for enhancing the generalization and reasoning capabilities of vision-language models in traffic incident understanding.