Scholarlink

<p>Accurate origin–destination (OD) matrix inference in urban traffic networks is difficult because of the challenges of limited sensor coverage and incomplete observations. To address this challenge, this study introduces a hybrid generative–analytical framework that integrates a generative adversarial network with a path flow proportion method to support OD inference by utilising partial intersection turning movement data. In this framework, the generative adversarial network learns statistically uniform path-choice proportions from sparse turning flow observations, which are subsequently translated into OD demand estimates through conditional inverse matrix operations. The proposed framework was evaluated on a simulated urban network at varying temporal aggregation levels and under different sensor coverage conditions. The results indicated that with moderate data constraints, the framework resulted in lower estimation errors compared with classical tomography-based estimators under identical observation conditions. As the observation windows shortened and sensor availability decreased, the accuracy of estimation progressively decreased, revealing observability limits inherent to turning-flow-based OD inference. By explicitly characterising these applicability boundaries, this study reveals how generative learning can improve data efficiency and mitigate error amplification under constrained but realistic sensing conditions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid Generative–Analytical Framework for Origin–Destination Matrix Inference Through Limited Turning Flow Observations

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Kuo Chi Yen,
  • Tsung Han Wang,
  • Chung Yung Wang

摘要

Accurate origin–destination (OD) matrix inference in urban traffic networks is difficult because of the challenges of limited sensor coverage and incomplete observations. To address this challenge, this study introduces a hybrid generative–analytical framework that integrates a generative adversarial network with a path flow proportion method to support OD inference by utilising partial intersection turning movement data. In this framework, the generative adversarial network learns statistically uniform path-choice proportions from sparse turning flow observations, which are subsequently translated into OD demand estimates through conditional inverse matrix operations. The proposed framework was evaluated on a simulated urban network at varying temporal aggregation levels and under different sensor coverage conditions. The results indicated that with moderate data constraints, the framework resulted in lower estimation errors compared with classical tomography-based estimators under identical observation conditions. As the observation windows shortened and sensor availability decreased, the accuracy of estimation progressively decreased, revealing observability limits inherent to turning-flow-based OD inference. By explicitly characterising these applicability boundaries, this study reveals how generative learning can improve data efficiency and mitigate error amplification under constrained but realistic sensing conditions.