BROWSE
Detail
Conference Paper
UniTraffic: Evidence-Centric Agentic Video Reasoning Across Traffic Anomalies, Violations, and Intent
Cited - time in 
Share 
Share
- Authors
- Jianxu Shangguan, Sangwon Kim, Hsiang-Wei Huang, Wei-Chieh (Winston) Sun, Junbin Lu, Pyongkun Kim, Kwang-Ju Kim, Chung-I Huang, Jenq-Neng Hwang
- Issue Date
- 2026-09
- Citation
- European Conference on Computer Vision (ECCV) 2026 Workshop : AI City Challenge, pp.1-13
- Publisher
- AI City Challenge
- Language
- English
- Type
- Conference Paper
- Abstract
- Traffic-safety video understanding spans heterogeneous tasks, from anomaly verification and structured violation reporting to pedestrian-intent reasoning, that are usually solved by separate per-task models. We argue that these tasks are different queries over the same latent traffic event, and present UniTraffic, an evidence-centric agentic framework that perceives a video once into a shared, timestamped Traffic Evidence Graph and then queries and renders that state per task. A coordinator plans the evidence each query needs; a two-pass policy combines a sparse global scan with query-conditioned temporal zoom; a fixed pool of cross-domain agents populates the graph; and an independent critic verifies claims against timestamped evidence before deterministic renderers emit each task’s output. Building on this design, our UWIPL_ETRI submissions ranked first on two out-of-distribution evaluations of the AI City Challenge 2026 Track 3, Traffic Violation Understanding (FETV, Track 7) and Pedestrian Situated Intent Visual Question Answering (PSI-VQA, Track 8), and achieved a competitive result on the in-domain Traffic Anomaly Reasoning (TAR) evaluation. Code: https://github.com/jxb1st/aicity2026-track3-7-8.
- Keyword
- Traffic-safety video understanding, Agentic reasoning, Evidence graph, Video question answering, Vision-language models, AI City Challenge
- KSP Keywords
- Cross-Domain, Evidence graph, Heterogeneous Tasks, Language Models, Task models, Traffic anomaly, Traffic violation, Video understanding, Visual question answering