A Tagged Traffic Accident Dataset for Machine Learning

This dataset contains tagged accident data and is provided for reproducibility for our journal paper Pablo Moriano, Andreas Berres, Haowen Xu, Jibonananda Sanyal. “Spatiotemporal Features of Traffic Help Reduce Automatic Accident Detection Time.” Expert Systems with Applications 244 (2024): 122813. https://doi.org/10.1016/j.eswa.2023.122813 The accompanying Data in Brief publication discusses the methodology behind the creation of these data. Berres, Andreas, Pablo Moriano, Haowen Xu, Sarah Tennille, Lee Smith, Jonathan Storey, and Jibonananda Sanyal. "A Traffic Accident Dataset for Chattanooga, Tennessee." Data in Brief (2024): 110675. The zip folder annotatedData.zip contains two subfolders: allData and bestData. The bestData folder contains all data for which a full neighborhood of five sensors upstream and five sensors downstream is available, whereas allData includes everything from bestData as well as data with a smaller number of neighboring sensors. Each folder contains one subfolder called accidents and one subfolder called non-accidents. The accidents folder contains one file per accident. The non-accidents folder contains files for the same location, day of the week and time as a corresponding accident, for each week during which there was no accident impact on the traffic. The file names in both folders are formatted as follows: yyyy-mm-dd-hhmm-rrrrrXaaa.a.csv, consisting of date (yyyy-mm-dd), time (hhmm in 24-hour format), and sensor name (rrrrrXaaa.a), which consists of road name (rrrrr; 5 alphanumerical characters), heading (X), and mile marker (aaa.a). For example, the file 2020-11-03-1611-00I24W182.8.csv contains data for an accident which occurred at 4:11 p.m. on November 3, 2020 on I-24 Westbound near the radar sensor at mile marker 182.8. The content of each CSV file is a timeseries of radar data beginning 15 minutes prior to the reported incident and ending 15 minutes after the reported incident. It also contains metadata, such as the accident type, etc. Each CSV file contains the following columns: incident at sensor(i): 1 for yes (accidents folder), 0 for no (non-accidents folder) road: road name with heading, e.g. 00I24E mile: mile marker of nearest radar sensor, e.g. 182.8 type: accident type, e.g. “Prop Damage (over)” for property damage exceeding a certain threshold. For non-accidents, the type is given as “None”. date: date of the data sample. For accidents, this is the date on which the accident occurred. For non-accidents, this is the date for which the non-accident data sample is collected. incident_time: time the reference accident was reported in hh:mm. This is the time which is provided in E-TRIMS as the time the 911 call was made. incident_hour: just the hour from the incident_time, in integer format. data_time: timestamp for the timeseries contained in the file in hh:mm:ss format. The timeseries consists of 30 second timesteps. weather: weather during data_time, based on data collected from NASA POWER. We used dry bulb temperature (°C), precipitation (mm/h), and wind speed (m/s) from the raw NASA POWER data to produce the classifications of rain (at least 1mm precipitation and temperatures above 2°C), snow (at least 1mm precipitation and temperatures at or below 2°C), and wind (wind speeds over 30 mph or 13.5 m/s). If there were no inclement weather conditions, we set the category to “--". light: light conditions during data_time. To produce this field, we collected sunrise, sunset, civil twilight start and civil twilight end times from https://sunrise-sunset.org, and derived the categories dawn, daylight, dusk, and dark using these start and end times. The last 33 columns contain radar data for the 11 sensors surrounding the accident or non-accident. For each sensor, we collected speed (mean over 30-second interval in miles per hour, or empty if no vehicles passed), volume (count of all vehicles passing during 30-second interval), and occupancy (mean % of occupancy over 30-second interval). These three variables are grouped in triples, of speed (k), volume (k), occupancy (k), where k indicates the sensor number relative to the closest sensor i to the incident, k<i indicate upstream sensors and k>i indicate downstream sensors. For example, speed (i-5) refers to the mean speed at the sensor which is 5 hops upstream from the accident, and volume(i+1) refers to the number of vehicles at the sensor immediately downstream from the accident. The folder metaData.zip contains the following files: Accidents.csv: cleaned-up accidents file with all accidents which happened on Chattanooga area highways between November 1, 2020 and April 29, 2021. We have removed accidents which happened on non-highway roads, and we have corrected the timestamps (which were in 12-hour format but missing a.m./p.m. markers) by cross-referencing light and weather conditions. WeatherDict.json: a dictionary containing the weather data synthesized from NASA POWER. LightDict.json: a dictionary containing the light data synthesized from Sunrise-and-Sunset. SensorTopology.csv: neighborhood information for each radar sensor in the Chattanooga area. SensorZones.geojson: polygons used to determine the nearest radar sensor for each accident location. Each polygon is tagged with the corresponding radar sensor’s name. Acknowledgments The authors would like to thank all parties who provided the datasets this work is based on: The accident and traffic data were provided by Tennessee Department of Transportation. These weather data was obtained from National Aeronautics and Space Administration (NASA) Langley Research Center's Prediction Of Worldwide Energy Resources (POWER) project funded through the NASA Earth Science Division. Sunrise and sunset times were obtained from https://sunrise-sunset.org. The authors would furthermore like to thank the Tennessee Department of Transportation and the Chattanooga Department of Transportation for their continued partnership and guidance. Finally, the authors would like to thank the U.S. Department of Energy (DOE) Vehicle Technologies Office for funding this work. This dataset has been prepared in part by the National Renewable Energy Laboratory, operated by Alliance for Sustainable Energy, LLC, for the DOE under Contract No. DE-AC36-08GO28308 and in part by UT-Battelle, LLC, under contract DE-AC05-00OR22725 with the US Department of Energy (DOE). The views expressed in the article do not necessarily represent the views of the DOE or the U.S. Government. The U.S. Government retains and the publisher, by accepting the article for publication, acknowledges that the U.S. Government retains a nonexclusive, paid-up, irrevocable, worldwide license to publish or reproduce the published form of this work, or allow others to do so, for U.S. Government purposes. DOE will provide public access to these results of federally sponsored research in accordance with the DOE Public Access Plan (http://energy.gov/downloads/doe-public-access-plan).本数据集包含带标注的事故数据（tagged accident data），旨在为我们发表于期刊的论文提供可复现性支持： Pablo Moriano, Andreas Berres, Haowen Xu, Jibonananda Sanyal. "Spatiotemporal Features of Traffic Help Reduce Automatic Accident Detection Time." 《专家系统与应用（Expert Systems with Applications）》244 (2024): 122813. https://doi.org/10.1016/j.eswa.2023.122813 配套的《数据简报（Data in Brief）》论文阐述了本数据集的构建方法： Berres, Andreas, Pablo Moriano, Haowen Xu, Sarah Tennille, Lee Smith, Jonathan Storey, and Jibonananda Sanyal. "A Traffic Accident Dataset for Chattanooga, Tennessee." 《数据简报（Data in Brief）》(2024): 110675. 压缩包annotatedData.zip包含两个子文件夹：allData与bestData。其中bestData文件夹收录了上下游各5个传感器（sensor）完整邻域范围内的所有数据；allData文件夹则涵盖bestData的全部数据，以及邻域内传感器数量更少的数据集。每个子文件夹下均设有accidents和non-accidents两个子目录：accidents目录下每个事故对应一个数据文件；non-accidents目录下的文件则与对应事故处于相同地点、星期和时段，且选取该周内未受事故交通影响的时段数据。 两个文件夹内的文件均遵循如下命名规则：yyyy-mm-dd-hhmm-rrrrrXaaa.a.csv，其组成部分依次为日期（格式为yyyy-mm-dd）、时间（24小时制的hhmm格式）以及传感器名称（格式为rrrrrXaaa.a）。其中，rrrrr代表5位字母数字组合的道路名称，X代表行驶方向，aaa.a代表里程标。例如，文件2020-11-03-1611-00I24W182.8.csv包含2020年11月3日下午4点11分，发生于I-24西向182.8英里处雷达传感器（radar sensor）附近的事故相关数据。 每个CSV文件的内容为报告事故前后各15分钟的雷达时序（timeseries）数据，同时附带事故类型等元数据。每个CSV文件包含以下列： 1. incident at sensor(i)：取值为1时表示该文件属于事故目录（accidents文件夹），取值为0时对应非事故目录（non-accidents文件夹） 2. road：带行驶方向的道路名称，例如00I24E 3. mile：最近雷达传感器的里程标，例如182.8 4. type：事故类型，例如"Prop Damage (over)"代表损失超过阈值的财产损失；非事故样本的该字段取值为"None" 5. date：数据样本的采集日期。对于事故样本，该日期为事故发生日期；对于非事故样本，该日期为非事故数据的采集日期 6. incident_time：参考事故的上报时间，格式为hh:mm。该时间取自E-TRIMS系统中911报警电话的拨打时间 7. incident_hour：从incident_time中提取的小时数，以整数形式呈现 8. data_time：文件内时序数据的时间戳，格式为hh:mm:ss，时序采样间隔为30秒 9. weather：data_time时刻的天气状况，基于NASA POWER项目的观测数据计算得到。我们利用原始NASA POWER数据中的干球温度（单位：℃）、降水量（单位：mm/h）和风速（单位：m/s）进行天气分类：降雨（降水量≥1mm且温度＞2℃）、降雪（降水量≥1mm且温度≤2℃）、大风（风速＞30mph即13.5m/s）；若无恶劣天气，则该字段取值为"--" 10. light：data_time时刻的光照条件。我们从https://sunrise-sunset.org获取日出、日落、民用暮光开始和结束时间，并据此将光照条件划分为黎明（dawn）、白天（daylight）、黄昏（dusk）和黑夜（dark）四类 文件最后33列为事故或非事故样本周边11个雷达传感器的观测数据。每个传感器对应3个变量：速度（30秒区间内的平均车速，单位：mph，无车辆通过时为空值）、车流量（30秒区间内通过的总车辆数）和占有率（30秒区间内的平均占有率百分比）。这三个变量以三元组形式分组，即speed(k)、volume(k)、occupancy(k)，其中k代表相对于事故最近传感器i的相对编号：k<i表示上游传感器，k>i表示下游传感器。例如，speed(i-5)代表事故上游第5个传感器的平均车速，volume(i+1)代表事故下游紧邻传感器的车流量。 压缩包metaData.zip包含以下文件： 1. Accidents.csv：经过清洗的事故数据集，涵盖2020年11月1日至2021年4月29日期间查塔努加地区高速公路上发生的全部事故。我们已剔除非高速公路上发生的事故，并通过交叉比对光照和天气条件，修正了原时间戳（原格式为12小时制但缺失上午/下午标记）的问题。 2. WeatherDict.json：存储基于NASA POWER数据合成的天气数据的字典文件。 3. LightDict.json：存储基于日出日落数据合成的光照数据的字典文件。 4. SensorTopology.csv：查塔努加地区各雷达传感器的邻域信息文件。 5. SensorZones.geojson：用于确定每个事故位置最近雷达传感器的多边形区域文件，每个多边形均标注了对应的雷达传感器名称。