VTTI: SHRP 2 Phase I Efforts - Standardization and Export of Training Dataset – Development of Data Repository

The purpose of this dataset is not to conduct research but to help researchers understand the driving related variables that are available in the second Strategic Highway Research Program (SHRP2) Naturalistic Driving Study (NDS). The data can also be useful for algorithm development and testing prior to running algorithms over all or part of the SHRP2 NDS dataset. The dataset was designed with the purpose of not requiring institutional review board (IRB) approval, and includes personally identifying information, including face video, and start and stop locations that can be openly shared. For this reason it cannot be used for research and should be thought of only as a training dataset. It resides on the InSight website and is available for download at no cost to qualified users. The focus of this effort was the time series data and video collected from the data acquisition system (DAS) that resides on the vehicle, including radar. Three vehicles from the Virginia Tech Transportation Institute (VTTI) fleet were driven from Blacksburg, VA to Durham, North Carolina to collect data on the same roads included in the roadway information database (RID). Data were collected for one week and include approximately 100 trips and more than 2,500 vehicle miles. Data were processed in the same way as for the original SHRP2 NDS including: the video processing, map matching to the RID, and the post processing done on the radar. It is downloadable in a 10 Hertz synchronized format. Access to the video is provided with each individual camera view separated and processed in the same manner as the main SHRP2 NDS dataset. The map matching yielded a heat map showing the trips driven on each road including origin and destination (i.e., does not require the current PII protection placed on the SHRP2 NDS). Therefore each trip is shown on the heat map from beginning to end. Since this was a limited data collection effort in which finding a naturally occurring scenario of interest was unlikely, several common events of interest were staged. The time of each staged scenario was documented. The following scenarios were included: Hard braking, Short headway, Low time to collision, High lateral acceleration, Vehicle cut in, and Flying pass.

本数据集的研发目的并非用于学术研究，而是协助研究者了解第二代战略公路研究计划（Strategic Highway Research Program, SHRP2）自然驾驶研究（Naturalistic Driving Study, NDS）中可获取的驾驶相关变量。该数据还可用于在针对SHRP2 NDS全量或子集数据集运行算法前，开展算法的开发与测试工作。本数据集的设计初衷为无需获得机构审查委员会（Institutional Review Board, IRB）的审批，其中包含个人身份识别信息，例如面部视频，以及可公开共享的行程起止位置数据。正因如此，本数据集无法用于学术研究，仅可作为训练数据集使用。它托管于InSight网站，面向符合资质的用户免费开放下载。

本次工作的核心聚焦于车载数据采集系统（Data Acquisition System, DAS，含雷达模块）所采集的时序数据与视频。研究团队从弗吉尼亚理工大学运输研究所（Virginia Tech Transportation Institute, VTTI）的车队中选取3台车辆，从弗吉尼亚州布莱克斯堡行驶至北卡罗来纳州达勒姆，在道路信息数据库（Roadway Information Database, RID）包含的相同路段上采集数据。本次数据采集周期为一周，涵盖约100次行程，总行驶里程超过2500车辆英里。

数据处理流程与原始SHRP2 NDS数据集保持一致，包括视频处理、与RID匹配的地图匹配，以及雷达数据后处理环节。该数据集以10赫兹同步格式提供下载。视频部分提供各独立摄像头视角的下载权限，其处理方式与SHRP2 NDS主数据集完全相同。地图匹配结果可生成热力图，展示各路段的行驶行程及起止点（无需像SHRP2 NDS那样采用当前的个人信息保护措施），因此热力图可完整呈现每次行程的始末轨迹。

由于本次数据采集规模有限，自然发生的目标场景较为稀缺，因此我们人工设置了多种常见的目标事件。各设置场景的时间均已记录在册，涵盖的场景包括：紧急制动（Hard braking）、跟车时距过短（Short headway）、碰撞时间阈值过低（Low time to collision）、横向加速度过高（High lateral acceleration）、车辆切入（Vehicle cut in）以及快速超车（Flying pass）。

搜集汇总

背景与挑战

背景概述

该数据集是VTTI为SHRP2项目开发的训练数据集，旨在帮助研究人员理解自然驾驶研究中的变量，并用于算法测试，但不能用于正式研究，因为它包含个人身份信息如面部视频和行程起止点。数据收集自三辆车一周的行程，覆盖约100次行程和超过2500英里，以10赫兹同步格式提供，并经过与SHRP2 NDS相同的标准化处理，包括视频分离、地图匹配和雷达后处理。数据集还模拟了多种常见驾驶场景（如急刹车、车辆切入等），以增强训练实用性，且热图显示完整行程路径，无需个人身份信息保护。