UniTraffic-Agent: AI Solution for Traffic Video Reasoning in AI City Challenge 2026
Researchers at MR-CAS have introduced UniTraffic-Agent, a novel system aimed at improving the comprehension of traffic videos for the 10th AI City Challenge Track 3. This system tackles the complexities of elucidating traffic incidents, their origins, and the timing of interactions—an area where multimodal large language models (MLLMs) struggle due to infrequent events and diverse perspectives in traffic footage. UniTraffic-Agent employs an observe-reason-act-verify methodology, extracting timestamped visual data and addressing all inquiries from a single clip in one go. It is specifically designed for Traffic Anomaly Reasoning (TAR) and includes evaluations for FETV and PSI-VQA. The findings are detailed in a paper available on arXiv (2608.13031), proposed as a solution for the AI City Challenge 2026.
Key facts
- UniTraffic-Agent is the MR-CAS solution for Track 3 of the 10th AI City Challenge.
- The challenge includes Traffic Anomaly Reasoning (TAR) and two out-of-domain evaluations: FETV and PSI-VQA.
- The system uses an observe-reason-act-verify workflow.
- It samples timestamped visual evidence and reasons over all questions from the same clip in one request.
- The goal is to explain how traffic events develop, why they happen, and when interactions occur.
- Traffic videos contain sparse events and varied viewpoints, posing challenges for MLLMs.
- The paper is available on arXiv with ID 2608.13031.
- The system is designed for intelligent transportation applications.
Entities
Institutions
- MR-CAS
- AI City Challenge