| 基于多任务迁移学习的交通警情信息自动处理方法 |
| |
| 引用本文： | 李昀轩,李萌,陆建,顾欣,郭娅明.基于多任务迁移学习的交通警情信息自动处理方法[J].中国公路学报,2022,35(9):1-12. |
| |
| 作者姓名： | 李昀轩 李萌 陆建 顾欣 郭娅明 |
| |
| 作者单位： | 1. 清华大学 土木工程系, 北京 100084;2. 东南大学 交通学院, 江苏 南京 211189;3. 北京工业大学 北京市交通工程重点实验室, 北京 100124 |
| |
| 基金项目： | 中国博士后科学基金项目(2021M701899) |
| |
| 摘 要： | 从交通警情数据中自动获取信息对于快速处理交通事故和提高交通管理水平具有重要的意义。为此，提出了一种基于多任务迁移学习的交通警情信息自动处理方法，该方法上游采用文本预训练模型作为共享参数层，下游建立多任务并行学习方法，实现对交通警情中的关键信息、类型和语义自动处理。选取江苏省苏州市2年内共120 191条原始交通警情作为试验数据，通过自动处理方法构建了一套标准的交通警情信息数据库。试验结果表明：所建立的关键信息抽取方法可以更精准地提取警情数据中的时间、地址和车牌信息；交通警情分类模型性能优于现有的深度学习模型，分类准确率达93%；基于局部特征增强的警情语义分析方法重点识别了警情中事故的严重程度和救援需求，识别准确率达87%。研究结论显示交通警情自动化处理方法具有良好的可移植性和实用性。
|
| 关 键 词： | 交通工程 交通警情 迁移学习 自然语言处理 交通事故 |
| 收稿时间： | 2022-01-18 |
An Auto-processing Method of Traffic Safety Information Based on a Multi-task Transfer Learning Algorithm |
| |
| LI Yun-xuan,LI Meng,LU Jian,GU Xin,GUO Ya-ming.An Auto-processing Method of Traffic Safety Information Based on a Multi-task Transfer Learning Algorithm[J].China Journal of Highway and Transport,2022,35(9):1-12. |
| |
| Authors: | LI Yun-xuan LI Meng LU Jian GU Xin GUO Ya-ming |
| |
| Affiliation: | 1. Department of Civil Engineering, Tsinghua University, Beijing 100084, China;2. School of Transportation, Southeast University, Nanjing 211189, Jiangsu, China;3. Beijing Key Laboratory of Traffic Engineering, Beijing University of Technology, Beijing 100124, China |
| |
| Abstract: | Auto-extraction of traffic safety information from traffic alarm reception data is of great significance for handling traffic crashes and improving traffic management. This paper established an auto-processing method based on a multi-task transfer learning algorithm that includes a text pre-training model as a shared parameter layer upstream, and a multi-task parallel processing method to automatically extract traffic safety information, types, and semantics. A total of 120 191 traffic alarm reception data items were collected over two years from a traffic control center in Suzhou Jiangsu, and a standard traffic safety information database was constructed with them. The experimental results show that the key information extraction method developed in this study can better extract the time, address, and license plate information from traffic alarm reception data. The performance of traffic crash classification achieved with the method proposed in this study is better than that of the conventional deep learning model: the classification accuracy is 93%. The traffic-safety semantic analysis method is based on local feature enhancement and focus on identifying the severity of crashes and rescue demands; its recognition accuracy is 87%. The results also demonstrate that the auto-processing method of traffic safety information proposed in this study has great portability and practicality. |
| |
| Keywords: | traffic engineering traffic alarm reception transfer learning algorithm natural language processing traffic crash |
|
| 点击此处可从《中国公路学报》浏览原始摘要信息 |
|
点击此处可从《中国公路学报》下载全文 |
|
