交通场站事件发现数据集

该数据集涵盖了17个关键交通路口与路段的监控视角，数据总量超1.1TB，包含超过30天的连续监控录像，覆盖了从清晨到深夜的全时段交通场景。参照百度地图Web-API交通态势评价标准，构建了细粒度的交通拥堵等级分类体系，将视频场景划分为：畅通、缓行、拥堵、严重拥堵4个核心类别，旨在为智能交通系统的拥堵检测与事件发现提供高质量的基准数据。具体采集过程分为数据获取与数据标注两个阶段。第一阶段为多源数据获取：首先，在地理空间上选取了包含国省干道（如G103京滨线）、城市主干路（如通马路、九周路）等在内的17个具有代表性的交通场站监测点位；其次，在时间维度上，以本地存储为载体，同步采集了长达一个多月的监控视频流（格式为hik/jvc/ntv，切片长度为10分钟）以及对应时段的小时级车辆通过量统计表（Excel格式），确保了非结构化视频数据与结构化流量数据的时空同步。第二阶段为数据清洗与半自动化标注：为了保证标注的客观性与科学性，摒弃了传统的主观目测标注，采用“基于多模态数据对齐的统计映射”方案。首先，对视频数据进行清洗，剔除缺失或损坏的片段；其次，依据百度Web-API的交通拥堵指数计算逻辑，结合每个监测点位的实际道路承载能力，利用分位数统计法计算流量阈值，将小时级流量数据映射为“畅通、缓行、拥堵、严重拥堵”四个等级；最后，通过编写自动化匹配算法，将清洗后的标签与对应时间戳的视频片段进行关联，生成最终的标注文件（CSV格式）。在标注完成后，由人工抽检员对极端天气或特殊时段的样本进行二次校验，确保标签与视觉感知的一致性。

This dataset covers surveillance views of 17 key traffic intersections and road sections, with a total data volume exceeding 1.1 TB, including over 30 consecutive days of surveillance footage that covers all-time traffic scenarios from early morning to late night. Referring to the traffic situation evaluation standard of Baidu Maps Web-API, a fine-grained traffic congestion level classification system was constructed, dividing video scenarios into four core categories: Free-flow, Slow-moving, Congested, and Severely Congested, aiming to provide high-quality benchmark data for congestion detection and incident discovery in intelligent transportation systems (ITS).

The specific collection process is divided into two stages: data acquisition and data annotation.

The first stage is multi-source data acquisition: First, 17 representative traffic station monitoring points were selected geographically, including national and provincial trunk roads (e.g., G103 Jingbin Line), urban arterial roads (e.g., Tongma Road, Jiuzhou Road), etc. Second, in terms of temporal dimension, using local storage as the carrier, synchronized collection of surveillance video streams spanning over one month (formatted as hik/jvc/ntv, with a clip length of 10 minutes) and hourly vehicle volume statistics tables (in Excel format) for corresponding periods was carried out, ensuring spatiotemporal synchronization between unstructured video data and structured traffic flow data.

The second stage is data cleaning and semi-automated annotation: To ensure the objectivity and scientificity of annotation, the traditional subjective visual annotation method was abandoned, and a "statistical mapping based on multimodal data alignment" scheme was adopted. First, the video data was cleaned to remove missing or corrupted clips; second, based on the traffic congestion index calculation logic of Baidu Web-API, combined with the actual road carrying capacity of each monitoring point, the quantile statistical method was used to calculate the flow threshold, and the hourly flow data was mapped to the four levels: Free-flow, Slow-moving, Congested, and Severely Congested; finally, by writing an automated matching algorithm, the cleaned labels were associated with the video clips of corresponding timestamps to generate the final annotation file (in CSV format). After annotation was completed, manual inspectors conducted secondary verification on samples from extreme weather or special periods to ensure consistency between the labels and visual perception.

搜集汇总

背景与挑战

背景概述

该数据集是一个专注于交通拥堵检测与事件发现的高质量基准数据集，涵盖17个关键交通路口与路段超过30天的连续监控录像，数据总量超1.1TB，覆盖全时段场景。它基于百度地图Web-API标准构建了细粒度的交通拥堵等级分类体系（畅通、缓行、拥堵、严重拥堵），并通过多源数据同步采集和半自动化标注方案，确保了视频数据与流量数据的时空对齐及标签准确性，旨在支持智能交通系统的研发与应用。