Roundabout-TAU
收藏资源简介:
Roundabout-TAU是由巴斯大学、华盛顿大学联合卡梅尔市交通部门构建的首个环岛交通异常理解基准数据集,包含342段真实路侧监控视频,覆盖28个摄像头视角的多样化交通场景。该数据集通过人工与GPT-5协同标注了2,064组多维度问答对,涵盖环境感知、物体定位、事件描述等细粒度语义信息,重点捕捉环岛场景中车辆违规、碰撞风险等复杂交互行为。其构建过程采用四阶段混合标注流程,严格保证数据质量,旨在推动智能交通系统中实时异常分析与语义理解的研究,为解决路侧监控视频的细粒度推理任务提供关键支持。
Roundabout-TAU is the first benchmark dataset for roundabout traffic anomaly understanding, jointly developed by the University of Bath, the University of Washington, and the Carmel Department of Transportation. It comprises 342 real roadside surveillance videos spanning diverse traffic scenarios from 28 distinct camera perspectives. A total of 2,064 multi-dimensional question-answer (QA) pairs were collaboratively annotated by human annotators and GPT-5, encompassing fine-grained semantic information including environmental perception, object localization, event description and more. The dataset specifically focuses on capturing complex interactive behaviors such as vehicle violations and collision risks in roundabout scenarios. A four-stage hybrid annotation pipeline was adopted during its construction to rigorously ensure data quality. It aims to advance research on real-time anomaly analysis and semantic understanding in intelligent transportation systems, and provide critical support for addressing fine-grained reasoning tasks based on roadside surveillance videos.
TAU-R1 数据集概述
数据集名称
Roundabout-TAU
核心目标
TAU-R1是一个基于视觉语言模型(VLM)的两层分层框架,旨在解决真实世界环岛场景下的交通异常理解任务。
数据集获取
数据集可通过以下地址获取:
- Hugging Face 仓库地址:https://huggingface.co/datasets/yl4300/Roundabout-TAU/tree/main
相关论文
- 论文标题:TAU-R1: Visual Language Model for Traffic Anomaly Understanding
- arXiv 链接:https://arxiv.org/abs/2603.19098
- arXiv ID:2603.19098
- 年份:2026
- 主要分类:cs.CV
项目代码
项目代码将在后续发布。
许可协议
本项目采用 MIT 许可证。




