Scholarlink

<p>This study develops a hybrid empirical–AI framework for calibrating urban traffic fundamental diagrams (FDs) using real-time, multi-source data. High-resolution video analytics (YOLOv8 and Deep SORT) were integrated with smartphone-based GPS and accelerometer measurements to derive speed–flow–density relationships across thirty heterogeneous urban sites. Four classical models—Greenshields, Greenberg, Underwood, and Drake—were initially calibrated using nonlinear least-squares regression, followed by parameter refinement through Genetic Algorithm (GA) and Bayesian Optimization (BO) to improve convergence and reduce estimation bias. Among the analytical models, the optimized Underwood formulation exhibited strong performance (R<sup>2</sup> = 0.96, RMSE = 4.3 km/h) among the evaluated formulations, reflecting the suitability of exponential decay structures under mixed and non-lane-based traffic conditions. For predictive benchmarking, a Support Vector Regression (SVR) model was trained on the same dataset, achieving high predictive accuracy (R<sup>2</sup> = 0.97) under the evaluated dataset, albeit with limited interpretability compared to physics-based models. The hybrid GA–BO calibration improved parameter accuracy by up to approximately 25% relative to baseline estimates across the analyzed sites. The final field-calibrated parameters—free-flow speed (~ 65 km/h), critical density (~ 32 veh/km), and jam density (~ 145 veh/km)—offer realistic guidance for capacity estimation and level-of-service assessment in developing urban networks. Overall, the proposed low-cost and scalable methodology provides an effective pathway toward data-augmented, empirically validated, and theoretically consistent FD modeling within heterogeneous urban traffic environments.</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 calibration of urban traffic fundamental diagrams using multi-source real-time data

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月6日

查看全文
  • Laith D. Alomari

摘要

This study develops a hybrid empirical–AI framework for calibrating urban traffic fundamental diagrams (FDs) using real-time, multi-source data. High-resolution video analytics (YOLOv8 and Deep SORT) were integrated with smartphone-based GPS and accelerometer measurements to derive speed–flow–density relationships across thirty heterogeneous urban sites. Four classical models—Greenshields, Greenberg, Underwood, and Drake—were initially calibrated using nonlinear least-squares regression, followed by parameter refinement through Genetic Algorithm (GA) and Bayesian Optimization (BO) to improve convergence and reduce estimation bias. Among the analytical models, the optimized Underwood formulation exhibited strong performance (R2 = 0.96, RMSE = 4.3 km/h) among the evaluated formulations, reflecting the suitability of exponential decay structures under mixed and non-lane-based traffic conditions. For predictive benchmarking, a Support Vector Regression (SVR) model was trained on the same dataset, achieving high predictive accuracy (R2 = 0.97) under the evaluated dataset, albeit with limited interpretability compared to physics-based models. The hybrid GA–BO calibration improved parameter accuracy by up to approximately 25% relative to baseline estimates across the analyzed sites. The final field-calibrated parameters—free-flow speed (~ 65 km/h), critical density (~ 32 veh/km), and jam density (~ 145 veh/km)—offer realistic guidance for capacity estimation and level-of-service assessment in developing urban networks. Overall, the proposed low-cost and scalable methodology provides an effective pathway toward data-augmented, empirically validated, and theoretically consistent FD modeling within heterogeneous urban traffic environments.