Online Parameter Estimation Methods for Adaptive Cruise Control Systems

Yanbing Wang, George Gunter, Matthew Nice, Maria Laura Delle Monache, Daniel B. Work

TL;DR

This paper introduces two novel online methods for estimating the parameters of adaptive cruise control (ACC) vehicle models in real-time. These methods, a recursive least-squares (RLS) approach and a particle filter (PF), offer comparable accuracy to traditional offline methods but with significant speedups, making them suitable for streaming data and real-time applications.

What Problem Does This Paper Address?

As vehicles with advanced driver-assistance systems (ADAS) like Adaptive Cruise Control (ACC) become more prevalent, understanding their impact on traffic flow is crucial. To accurately model and predict this impact, we need precise car-following models that capture ACC vehicle behavior. However, traditional methods for estimating the parameters of these models are often computationally intensive and offline, meaning they require all data to be collected before analysis. This limits their utility for real-time applications or when dealing with continuously growing datasets. This paper addresses this gap by developing and evaluating two online parameter estimation methods that can operate in real-time.

Methods

The paper proposes and evaluates two online methods for estimating car-following model parameters:

  • Recursive Least-Squares (RLS): This method reformulates the car-following model into a linear state-space representation. It then uses a recursive least-squares approach to estimate the parameters of this linear system. The key advantage here is that it avoids solving differential equations repeatedly and can update parameter estimates incrementally as new data arrives.
  • Particle Filter (PF): This method tackles the problem by treating both the vehicle's state (speed, spacing) and the model parameters as part of an augmented state vector. It then employs a particle filter, a sequential Monte Carlo method, to estimate this augmented state. The PF is particularly useful for nonlinear systems and can handle noise in both the system dynamics and measurements.

Both online methods are compared against a Batch Optimization approach, which is a standard offline method that solves the car-following model as a constraint within an optimization problem, typically using a gradient-based approach.

Datasets

The paper utilizes two types of datasets for evaluation:

  • Synthetic Data: This data was generated by simulating a car-following model with known parameters. This allows for a direct assessment of the accuracy of each estimation method in recovering true parameters without the influence of sensor noise.
  • Empirical Data: Real-world data was collected from a 2019 model year ACC-equipped vehicle using its on-board radar system. This data includes measurements of speed, space gap, and relative velocity. The accuracy of the radar data was validated against simultaneous GPS measurements, showing good agreement.

Key Insights and Ideas

  • Real-time Parameter Estimation is Feasible: The paper demonstrates that accurate car-following model parameters can be estimated in real-time using both RLS and PF methods, opening doors for dynamic control and traffic simulation.
  • RLS Offers Significant Speedup: The RLS method provides a dramatic speed improvement over offline batch methods, making it highly practical for large datasets and real-time applications.
  • Particle Filter Handles Complexity: The PF, while slower than RLS, offers a robust approach for potentially nonlinear dynamics and provides a distribution of parameter estimates, which can be valuable for uncertainty quantification.
  • Model Identifiability is Crucial: The paper highlights the importance of parameter identifiability, noting that some parameters might not be uniquely determinable from equilibrium driving data alone.
  • ACC Models Can Be String Unstable: Consistent with prior research, the estimated models often exhibit string instability, suggesting that ACC systems, like human drivers, can contribute to phantom traffic jams.

Key Results

  • Accuracy: All three methods (Batch Optimization, RLS, and PF) achieved comparable low Mean Absolute Errors (MAE) in speed and spacing when evaluated on both synthetic and real data.
  • Runtime:
    • The RLS method was significantly faster, achieving runtimes of 0.0094 seconds on synthetic data and 0.055 seconds on real data. This is orders of magnitude faster than the batch method.
  • The Batch Optimization method took 19.3 seconds (synthetic) and 18.59 seconds (real data).
  • The Particle Filter was the slowest, taking 49 seconds (synthetic) and 73.5 seconds (real data). However, it still completed its computation within the duration of the experiment, making it a viable online method.
  • String Stability: All methods generally identified models that were likely string unstable, with the PF providing a probabilistic estimate (e.g., 98.52% likely string unstable on synthetic data).

Limitations and Future Work

  • Non-linear Dynamics: While the paper focuses on a linear car-following model, the authors acknowledge that real-world ACC behavior can be more complex and may involve non-linear dynamics.
  • Parameter Identifiability: The paper notes that certain parameters might not be uniquely identifiable from equilibrium driving data, suggesting the need for careful experimental design or more informative data.
  • Future Work: The authors propose to expand on the online estimation methods by:
  • Developing the RLS estimation problem into a real-time estimation method.
  • Investigating car-following model calibration techniques that leverage estimated acceleration data available from the vehicle's CAN bus, in addition to speed and spacing.

Summary updated

Abstract

Modeling Adaptive Cruise Control (ACC) vehicles enables the understanding of the impact of these vehicles on traffic flow. In this work, two online methods are used to provide real time system identification of ACC enabled vehicles. The first technique is a recursive least squares (RLS) approach, while the second method solves a nonlinear joint state and parameter estimation problem via particle filtering (PF). We provide a parameter identifiability analysis for both methods to analytically show that the model parameters are not identifiable using equilibrium driving. The accuracy and computational runtime of the online methods are compared to a commonly used offline simulation-based optimization (i.e., batch optimization) approach. The methods are tested on synthetic data as well as on empirical data collected directly from a 2019 model year ACC vehicle using data from sensors that are part of the stock ACC system. The online methods are scalable and provide comparable accuracy to the batch method. RLS runs in real time and is two orders of magnitude faster than the batch method for modest sized (e.g., 15 min) datasets. The particle filter also runs in real-time, and is also suitable in streaming applications in which the datasets can grow arbitrarily large.
BibTeX
@article{wang2021online,
  author = {Yanbing Wang and George Gunter and Matthew Nice and Maria Laura Delle Monache and Daniel B. Work},
  title = {{Online Parameter Estimation Methods for Adaptive Cruise Control Systems}},
  journal = {IEEE Trans. Intell. Veh.},
  year = {2021},
  url = {https://openreview.net/forum?id=86gSSKOZyZ},
}
Figure