Learning and Inferring a Driver's Braking Action in Car-Following Scenarios

Wenshuo Wang, Junqiang Xi, Ding Zhao

TL;DR

This paper introduces a novel GMM-HMM approach to predict a driver's braking intent in car-following scenarios, outperforming traditional SVM methods. By modeling the perception-decision-action loop using vehicle sensor data, this method offers a more robust and accurate way to enhance forward collision warning systems and improve traffic safety.

What Problem Does This Paper Address?

Rear-end collisions are a significant cause of accidents, particularly in car-following situations. Existing forward collision warning (FCW) systems, while beneficial, often suffer from high false-positive alarm rates, reducing their effectiveness and user acceptance. A key challenge is accurately predicting a driver's intent to brake before they actually do, allowing the system to issue timely warnings or interventions. Previous approaches have relied on foot gesture analysis or physiological signals, which have limitations in terms of practicality or accuracy. This paper aims to bridge this gap by developing a method that infers braking intent from readily available vehicle sensor data, considering the driver's historical behavior and the current driving context.

Methods

The core of this paper is the development of a Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) framework.

  • Perception-Decision-Action (PDA) Perspective: The model frames driver behavior as a sequence of perception, decision, and action.
  • Explanatory Variables: The driving situation is characterized by four key variables: time to collision (TTC), relative speed ($\Delta v$), relative distance ($L$), and the ego vehicle's speed ($v_E$).
  • GMM for Stochastic Relationships: A Gaussian Mixture Model (GMM) is used to model the complex, stochastic relationships between the driving situation variables and the driver's braking action. Each mixture component of the GMM represents a hidden mode.
  • HMM for Temporal Dynamics: A Hidden Markov Model (HMM) then uses these GMM-derived modes to infer the driver's braking action over time, capturing the sequential nature of driving behavior.
  • Binary Braking Action: The braking action is simplified to a binary output: 1 for braking and 0 for no braking.
  • Comparison Methods: The proposed GMM-HMM is compared against a standard Support Vector Machine (SVM) and an SVM with Bayesian Filtering (SVM-BF).

Datasets

The study utilizes a large-scale, naturalistic driving dataset from the University of Michigan Safety Pilot Model Deployment (SPMD) program. This dataset includes:

  • Data from 56 drivers over three years.
  • Over 72,000 car-following events.
  • Data collected at 10 Hz, including vehicle speed, relative speed, relative distance, road curvature, throttle opening, and brake pedal position, all recorded from onboard sensors.

Key Insights and Ideas

  • Beyond Foot Gestures: The paper demonstrates that inferring braking intent can be effectively achieved using vehicle sensor data (perception-decision-action) rather than relying on less practical methods like foot gesture analysis or physiological signals.
  • GMM-HMM Synergy: Combining GMM for modeling complex relationships and HMM for temporal dynamics proves to be a powerful approach for capturing the nuanced and stochastic nature of driver behavior.
  • Importance of Context: The model's success highlights the critical role of considering not just the immediate driving situation but also the driver's historical data and the sequential nature of their actions.
  • Trade-off in Critical Value: The choice of the critical value ($Br^c$) for decoding the model's output involves a trade-off between accuracy, sensitivity, and specificity, suggesting that optimal system tuning might be driver-dependent.

Key Results

The GMM-HMM method achieved superior performance compared to SVM and SVM-BF across all metrics:

  • Average Performance:
    • Accuracy: 89.41%
    • Sensitivity: 83.42%
    • Specificity: 97.41%
  • Best Performance: For some drivers, the GMM-HMM achieved accuracies as high as 96.42%, sensitivities of 95.41%, and specificities of 99.54%.
  • Comparison: The GMM-HMM improved accuracy by 26.37% over SVM and 5.05% over SVM-BF. Sensitivity saw even larger gains (39.06% over SVM, 8.03% over SVM-BF). Specificity also showed improvement.

Limitations and Future Work

  • Road Curvature: The current model does not explicitly account for the influence of large road curvatures, which can affect driver braking decisions. Future work should incorporate this factor.
  • Bounded Variables: The model did not explicitly consider the bounded nature of driving data (e.g., typical ranges for relative distance or speed). Developing a more robust Gaussian mixture model to fit such data is suggested.
  • Driving Styles: The paper focuses on inferring the intent to brake, not the style of braking (e.g., aggressive vs. gentle). Future research could extend this to include such stylistic elements.
  • Critical Value Tuning: The influence of the critical value ($Br^c$) on performance metrics suggests that personalized tuning might be beneficial for FCW systems.

Summary updated

Abstract

Accurately predicting and inferring a driver's decision to brake is critical for designing warning systems and avoiding collisions. In this paper we focus on predicting a driver's intent to brake in car-following scenarios from a perception-decision-action perspective according to his/her driving history. A learning-based inference method, using onboard data from CAN-Bus, radar and cameras as explanatory variables, is introduced to infer drivers' braking decisions by combining a Gaussian mixture model (GMM) with a hidden Markov model (HMM). The GMM is used to model stochastic relationships among variables, while the HMM is applied to infer drivers' braking actions based on the GMM. Real-case driving data from 49 drivers (more than three years' driving data per driver on average) have been collected from the University of Michigan Safety Pilot Model Deployment database. We compare the GMM-HMM method to a support vector machine (SVM) method and an SVM-Bayesian filtering method. The experimental results are evaluated by employing three performance metrics: accuracy, sensitivity, specificity. The comparison results show that the GMM-HMM obtains the best performance, with an accuracy of 90%, sensitivity of 84%, and specificity of 97%. Thus, we believe that this method has great potential for real-world active safety systems.
BibTeX
@article{wang2018learning,
  author = {Wenshuo Wang and Junqiang Xi and Ding Zhao},
  title = {{Learning and Inferring a Driver's Braking Action in Car-Following Scenarios}},
  journal = {IEEE Trans. Vehicular Technology},
  year = {2018},
  doi = {10.1109/tvt.2018.2793889},
}
Figure