Second-order models and traffic data from mobile sensors

Benedetto Piccoli, Ke Han, Terry L. Friesz, Tao Yao, Junqing TangTransportation 2015 · DOI · arXiv

Abstract

Traditional first-order traffic models often fail to capture the complex, non-equilibrium dynamics of congested highway traffic, leading to significant inaccuracies in estimating environmental impacts. This paper evaluates the efficacy of a second-order Phase Transition Model (PTM) in reconstructing high-order traffic quantities—such as acceleration and emissions—from sparse mobile sensor data. Using the NGSIM dataset, we demonstrate that while first-order models are sufficient for general travel time estimation, second-order models are essential for capturing the "stop-and-go" waves that drive fuel consumption and tailpipe emissions. Our findings highlight a critical sampling threshold of one second for acceleration accuracy and propose robust data fusion and correction techniques to mitigate errors in low-penetration data environments.

Introduction

Most contemporary traffic management systems rely on first-order macroscopic models, most notably the Lighthill-Whitham-Richards (LWR) model. These models operate under the assumption of a functional relationship between traffic density and speed, implying that traffic remains in a state of equilibrium. While computationally efficient and effective for estimating aggregate travel times, the LWR model fails to account for the instabilities inherent in congested traffic, such as "phantom jams" and stop-and-go oscillations.

As the focus of transportation research shifts toward sustainability, there is an increasing need to monitor higher-order quantities like vehicle acceleration, which are direct precursors to fuel consumption and emissions. These metrics are highly sensitive to micro-movements that first-order models ignore. This study investigates how the limitations of mobile sensor data—specifically low sampling frequencies and low penetration rates—impact our ability to monitor the environmental footprint of highway traffic, and proposes a second-order modeling framework to address these challenges.

Theoretical Framework: The Phase Transition Model (PTM)

To move beyond the limitations of equilibrium-based modeling, this research employs a second-order Phase Transition Model (PTM). Unlike first-order models that track only density ($\rho$), the PTM introduces a second state variable ($q$) representing the deviation from the equilibrium state.

This additional degree of freedom allows the model to capture the "scatter" observed in empirical traffic data. In real-world scenarios, two segments with identical traffic density may exhibit vastly different speeds due to local instabilities or driver reaction delays. The PTM treats traffic as a system capable of existing in different phases (e.g., free-flow vs. congested), allowing for a more authentic representation of the transitions between these states.

Methodology

Data Fusion Schemes

The study proposes three distinct methods for integrating Lagrangian (individual vehicle trajectory) data into the Eulerian (fixed-grid) PTM framework:

  1. Source-Term Integration (Methods 1 & 2): These methods utilize the Siebel-Mauser term to account for the relaxation time of drivers—the period it takes for traffic to settle back toward an equilibrium state after a perturbation.
  2. Conservation Law Approach (Method 3): This "no-source" approach treats the PTM as a pure system of conservation laws, solving a quadratic equation to derive hidden state variables from available sensor pings.

Robust Denoising and Differentiation

Raw GPS data is notoriously noisy. To derive meaningful acceleration values, the authors utilize Tikhonov regularization. This numerical differentiation technique ensures that calculated acceleration is physically consistent with the vehicle's trajectory over time. Unlike standard smoothing filters (such as RLOESS), which may erroneously categorize sudden, real speed changes as outliers, Tikhonov regularization preserves the integrity of the stop-and-go signals essential for emission modeling.

Experimental Setup

The research utilizes the Next Generation SIMulation (NGSIM) dataset, specifically focusing on the I-80 highway segment in Emeryville, California. The NGSIM data provides a "ground truth" with vehicle trajectories recorded at 0.1-second intervals. To test the robustness of the proposed models, this high-resolution data was artificially degraded to simulate realistic mobile sensor limitations:

  • Sampling Intervals: Down-sampled from 0.1s to 1s, 2s, and 3s.
  • Penetration Rates: Reduced from 100% to as low as 2% of the total vehicle population.

Results and Discussion

The Sampling Threshold

A key finding of the study is the existence of a "sampling threshold" for higher-order traffic quantities. While density and mean speed can be estimated with reasonable accuracy at 3-second intervals, the error in acceleration estimation increases exponentially as the interval grows. Specifically, the error in acceleration jumped from approximately 21% at a 1-second interval to over 63% at a 3-second interval.

Model Performance: PTM vs. LWR

The second-order PTM consistently outperformed the classical LWR model across all test cases. At 100% penetration, the PTM achieved a mean relative error for density of 8.7%, compared to 12.6% for the LWR model. The PTM’s ability to "see" the perturbations and instabilities within the traffic flow makes it significantly more reliable for congested highway scenarios.

Emission Robustness and Correction Factors

The study found that estimating emissions by aggregating data into spatial "bins" (an Eulerian approach) is more resilient to sampling errors than attempting to track individual vehicle emissions (a Lagrangian approach). The aggregation process acts as a natural filter for noise.

Furthermore, the authors identified a consistent trend where low sampling rates lead to an underestimation of total emissions. By applying a linear regression-based correction factor, the researchers were able to keep the error in Hydrocarbon (HC) emission estimates between 10-14%, even in scenarios with low penetration rates and 3-second sampling intervals. Without this correction, errors were nearly twice as high.

Conclusion

This research demonstrates that second-order models like the PTM are vital for accurate traffic monitoring in the era of green transportation. By accounting for non-equilibrium dynamics, the PTM provides a much clearer picture of the environmental impact of highway congestion than traditional models. While the current framework is optimized for highways, future work will focus on adapting these models for urban arterials, where traffic behavior is governed by signalized intersections rather than internal fluid dynamics. Additionally, integrating vehicle-specific characteristics into the PTM framework could further refine the accuracy of real-time emission management systems.

Summary updated

BibTeX
@inproceedings{piccoli2015second,
  author = {Benedetto Piccoli and Ke Han and Terry L. Friesz and Tao Yao and Junqing Tang},
  title = {{Second-order models and traffic data from mobile sensors}},
  booktitle = {Transportation},
  year = {2015},
  doi = {10.1016/j.trc.2014.12.013},
}

Loading related content...

Figure