Developing an Intelligent Transportation Management Center (ITMC) with a Safety Evaluation Focus for Smart Cities (04-110)

Project Description: The Intelligent Transportation Management Center (ITMC) project was launched to address the limitations of traditional Transportation Management Centers (TMCs) by integrating advanced technologies such as machine learning, big data science, and image processing. The research, spanning from mid-November to mid-December 2022, focused on a signalized intersection at H Street and Broadway in Chula Vista, San Diego. The ITMC aimed to go beyond the human-dependent operations of conventional TMCs, which are prone to errors, by automating the detection and analysis of near-crash events across multiple transportation modes including vehicles, trucks, bicycles, motorcycles, and pedestrians. The use of Post Encroachment Time (PET), a recognized surrogate safety measure (SSM), allowed for a more proactive approach to safety evaluation. To process the collected video files for detection and tracking of the road users, YOLOX, the latest iteration of the YOLO series, was used as the deep learning modeling framework. This project involved collecting vast amounts of video data to not only detect incidents but also to perform spatiotemporal analyses, thus identifying the conditions under which near- crashes are most likely to occur. By examining these incidents at various times—peak versus off- peak hours, and weekdays versus weekends—and during different stages of the traffic signal cycle, the research provides insights into the dynamics of safety risks at the intersection. The spatial analysis component of the study produced heatmap visualizations that pinpoint areas of frequent near-crash events, effectively highlighting the intersection’s safety hotspots. This visual tool aids in illustrating the concentration of risk exposure, which is essential for transportation planning and decision-making. Ultimately, the project's comprehensive approach offers a potential model for enhancing transportation safety through technology-driven solutions. Data Scope: Four high-definition cameras were installed by the research team on the signal mast arms at the intersection of H Street and Broadway in Chula Vista, San Diego. These cameras captured continuous 720p high-definition video at 10 frames per second, meticulously documenting the traffic conditions from all directions. The footage was systematically saved on an hourly basis in Audio Video Interleave (.avi) format, employing Motion JPEG (MJPEG) compression, culminating in a comprehensive dataset encompassing 3.8 million interactions among road users. This dataset was enriched with a variety of associated variables for each interaction, among which the PET variable was paramount for the safety evaluation. Utilizing this variable, a detailed analysis was conducted to assess the potential safety risks at the intersection. The intricate details of the dataset and the variables are further elaborated upon in the subsequent section of the documentation. Video files were processed using the YOLOX deep learning model, and the data management was facilitated by connecting each camera to an NVIDIA® Jetson AGX Xavier™ edge device, fitted with a 5TB USB hard drive. A Cradlepoint IBR1700 Series Ruggedized Router with LTE interface was employed to ensure the efficient retrieval of the video files over the cellular network, with the files being transferred to the university GPU server at notos.sdsu.edu for object detection and further analysis. Data Specification: The processed data was meticulously organized into structured formats compatible with SQL databases and CSV files, featuring 23 distinct variables. These variables are detailed in the table (included below), which includes the type of each variable and the range of values it encompasses.

项目概述：智能交通管理中心（Intelligent Transportation Management Center，ITMC）项目旨在弥补传统交通管理中心（Transportation Management Centers，TMCs）的不足，通过集成机器学习、大数据科学与图像处理等先进技术构建。本研究于2022年11月中旬至12月中旬开展，聚焦圣地亚哥丘拉维斯塔市H街与百老汇交叉口的信号控制路口。

传统TMC依赖人工操作且易出现差错，ITMC则突破该局限，实现对机动车、货车、自行车、摩托车及行人等多类交通参与者的近碰撞事件自动化检测与分析。研究采用冲突后时间（Post Encroachment Time，PET）这一公认的替代安全指标（Surrogate Safety Measure，SSM），为交通安全评估提供了更具前瞻性的分析路径。

为处理采集到的视频文件以实现交通参与者的检测与跟踪，本项目采用YOLO系列最新迭代模型YOLOX作为深度学习建模框架。本次项目通过采集海量视频数据，不仅完成事件检测，还开展时空分析，以此识别近碰撞事件的高发场景。研究通过对比高峰与平峰时段、工作日与周末的交通场景，以及交通信号周期的不同阶段，深入剖析该交叉口的安全风险动态。

本研究的空间分析模块生成热图可视化结果，可精准定位近碰撞事件的频发区域，有效识别该交叉口的安全热点。该可视化工具可直观展示风险暴露的集中程度，对于交通规划与决策制定具有重要价值。最终，本项目采用的综合技术方案为依托技术驱动的解决方案提升交通安全提供了可借鉴的范式。

数据范围：研究团队在该交叉口的信号杆臂上安装4台高清摄像机，以10帧每秒的帧率采集720p高清连续视频，全方位记录各方向的交通状况。视频素材按小时为单位存储，采用音频视频交错（Audio Video Interleave，.avi）格式封装，并使用运动JPEG（Motion JPEG，MJPEG）压缩算法，最终形成包含380万次交通参与者交互记录的完整数据集。数据集为每一次交互记录补充多维度关联变量，其中冲突后时间（PET）变量是交通安全评估的核心指标。基于该变量，研究团队开展精细化分析以评估该交叉口的潜在安全风险。数据集的详细构成与变量说明将在后续文档章节中进一步阐述。

视频文件通过YOLOX深度学习模型完成处理，数据管理环节通过将每台摄像机连接至搭载5TB USB硬盘的NVIDIA® Jetson AGX Xavier™边缘设备实现。同时采用Cradlepoint IBR1700系列加固型LTE路由器，依托蜂窝网络实现视频文件的高效回传，将文件传输至位于notos.sdsu.edu的圣地亚哥州立大学GPU服务器，用于目标检测与后续分析。

数据规范：经处理后的数据被严格整理为兼容SQL数据库与CSV文件的结构化格式，共包含23个独立变量。各变量的类型与取值范围详见下文表格。