作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2024, Vol. 50 ›› Issue (7): 333-341. doi: 10.19678/j.issn.1000-3428.0068215

• 开发研究与工程应用 • 上一篇    下一篇

基于线性低秩卷积与道路网络的城市流量推断

刘树林, 李红军*(), 甘雨金, 罗茜雅   

  1. 成都理工大学计算机与网络安全学院, 四川 成都 610059
  • 收稿日期:2023-08-10 出版日期:2024-07-15 发布日期:2023-12-19
  • 通讯作者: 李红军
  • 基金资助:
    国家自然科学基金(42050104); 成都理工大学自然资源部深时地理与环境重建及应用重点实验室开放基金(DGERA20221102)

Urban Flow Inference Based on Linear Low-Rank Convolution and Road Network

Shulin LIU, Hongjun LI*(), Yujin GAN, Xiya LUO   

  1. College of Computer Science and Cyber Security, Chengdu University of Technology, Chengdu 610059, Sichuan, China
  • Received:2023-08-10 Online:2024-07-15 Published:2023-12-19
  • Contact: Hongjun LI

摘要:

细粒度城市流量推断(FUFI)旨在从粗粒度交通流量中推断出真实的细粒度交通流量, 以代替在现实世界中大量传感器设备的作用。现有的FUFI方法仅考虑到时间、天气等外部因素特征, 忽略了道路网络特征对城市交通流的重要影响。此外, 现有方法使用的传统残差网络结构对交通流的低级特征捕获能力不足, 低级特征容易在网络深层消亡。为解决以上问题, 提出一种使用线性低秩卷积与全局注意力Transformer的细粒度城市流量推断模型LLCGAT, 以更好地捕获交通流的低级特征并融合道路网络特征的学习。该模型在考虑外部因素的基础上, 首先将城市的道路网络作为重要的特征与交通流特征融合, 并使用广泛激活的线性低秩卷积对综合特征进行特征提取, 然后将综合特征与道路网络特征分别接入注意力Transformer的编码器和解码器中以进一步捕获交通流的全局空间分布。在TaxiBJ-P1和XiAn两个真实世界数据集上的实验结果表明, LLCGAT模型将平均绝对百分比误差分别降低了3.3%和10.7%, 均方根误差分别降低了2.3%和2.4%, 平均绝对误差分别降低了3.8%和6.3%。

关键词: 智能交通系统, 细粒度城市流量推断, 道路网络特征, 线性低秩卷积, Transformer架构

Abstract:

Fine-grained Urban Flow Inference (FUFI) aims to elucidate real fine-grained traffic flows from coarse-grained traffic flows to replace a large number of sensor devices in the real world. Existing fine-grained urban flow inference methods only consider external factors such as the time of day and weather, ignoring the important impact of road network characteristics on urban traffic flows. In addition, the traditional residual network structure used by existing methods is insufficient to capture the low-level features of traffic flows, which tend to die out in the deeper layers of the network. To address these problems, a fine-grained urban traffic flow inference model based on Linear Low-rank Convolution with Global Attention Transformer (LLCGAT) is proposed to better capture the low-level features of traffic flows and incorporate the learning of road network features. The model combines the road network of a city as an important feature with traffic flow features, considering external factors, and uses a widely activated linear low-rank convolution to extract integrated features. Subsequently, these features are plugged into the encoder and decoder of the attention transformer to further capture the global spatial distribution of traffic flows. Experiments are conducted on two real-world datasets, namely, TaxiBJ-P1 and XiAn. The results indicate that the LLCGAT model reduces the mean absolute percentage error by 3.3% and 10.7%, the root-mean-square error by 2.3% and 2.4%, and the mean absolute error by 3.8% and 6.3%, respectively.

Key words: Intelligent Transportation System(ITS), Fine-grained Urban Flow Inference(FUFI), road network feature, linear low-rank convolution, Transformer architecture