Evaluation of Infrastructure-based Warning System on Driving Behaviors-A Roundabout Study

Cong Zhang, Chi Tian, Tianfang Han, Hang Li, Yiheng Feng, Yunfeng Chen, Robert W. Proctor, Jiansong ZhangarXiv 2023 · OpenReview

TL;DR

This paper introduces a novel co-simulation platform for evaluating infrastructure-based warning systems at roundabouts. Their study demonstrates that timely warnings significantly improve driving safety and comfort by reducing abrupt maneuvers, and they developed a personalized machine learning model to predict driver intent, achieving high accuracy.

What Problem Does This Paper Address?

Traffic crashes, particularly at intersections, remain a significant safety concern. Roundabouts, while designed to improve traffic flow, present unique challenges for drivers, leading to common crash types like "failed to yield." Existing research on driver assistance systems often focuses on general road conditions or simpler intersections. There's a need to specifically investigate how infrastructure-based warnings, delivered via V2X communication, can influence driver behavior in complex scenarios like roundabouts and how to predict driver responses to these warnings for more personalized safety systems.

Methods

The authors developed a customized co-simulation platform by integrating SUMO (Simulation of Urban Mobility) for traffic simulation and Webots for vehicle and driving simulation. This platform allows for realistic modeling of real-world environments, precise control of vehicle states, and seamless integration of V2X communication.

They conducted a driving simulator study using this platform, replicating a real-world roundabout in Ann Arbor, Michigan. Participants navigated the roundabout under varying conditions:

  • Aggressiveness levels of an "aggressive vehicle" (low, medium, high).
  • Warning levels (no warning, warning 1 second in advance, warning 2 seconds in advance).

To analyze driver behavior, they collected vehicle trajectory data (speed, acceleration) and eye-tracking data. Key metrics included:

  • Time-to-collision (TTC): A safety indicator of the time until a potential collision.
  • Crash Potential Index (CPI): Measures safety impact considering evasive actions.
  • Deceleration Rate to Avoid Crash (DRAC): Reflects required deceleration to avoid a crash.
  • Acceleration Noise (AN): Measures the variability in acceleration.
  • Pupil Diameter (PD): An indicator of cognitive load.
  • Fixation Duration: A measure of visual attention.

A significant contribution is the development of a personalized intention prediction model. This model uses six features (ego vehicle speed, time headway, AN, DRAC, mean fixation duration, average pupil diameter) to predict whether a driver will "stop or go" when a warning is issued. Several machine learning models were evaluated, with XGBoost showing the best performance.

Datasets

The study did not use pre-existing datasets for training or evaluation. Instead, it generated its own data through the driving simulator experiment with 36 participants. The "real-world roundabout in Ann Arbor, Michigan" served as the basis for the simulation environment.

Key Insights and Ideas

  • Timely warnings are crucial: Advanced warnings, especially those issued 2 seconds in advance, significantly improve driving safety by providing more time to react and reducing abrupt maneuvers.
  • Comfort and safety go hand-in-hand: Earlier warnings lead to smoother driving, reduced harsh decelerations, and increased driving comfort, not just improved objective safety metrics.
  • Personalized prediction is key: Drivers react differently to warnings. A personalized model that considers individual driving behaviors and visual attention (eye-tracking) is essential for effective warning systems.
  • Roundabouts are complex: The study highlights the inherent difficulty of roundabouts, especially when interacting with aggressive drivers, underscoring the need for advanced safety interventions.
  • "Rush out" behavior is dangerous: Drivers who disregard warnings and attempt to "rush out" face significantly lower TTCs, indicating a higher risk of immediate collision.

Key Results

  • Warning Effectiveness: Warnings significantly improved safety metrics (CPI, DRAC) compared to no-warning scenarios. A 2-second warning led to higher TTC and lower maximum deceleration compared to a 1-second warning.
  • Aggressiveness Impact: Higher aggressive levels of other vehicles led to lower TTC and higher DRAC, indicating more dangerous situations and more aggressive driver responses.
  • Risk Perception: Drivers perceived higher risk in more aggressive scenarios, but this perception didn't always directly translate to objective safety metrics (TTC, CPI) in counterintuitive ways, suggesting drivers might take more aggressive actions to compensate for perceived risk.
  • Intention Prediction: The XGBoost model achieved a prediction accuracy of 94.8%, with a precision of 95.56% and a recall of 97.73% for predicting driver stop-or-go decisions when a warning is issued.

Limitations and Future Work

  • Predefined Aggressive Behavior: The aggressive vehicle's behavior was predefined. Future work should focus on developing accurate prediction models for aggressive vehicle behavior in real-world implementations to determine if they will yield or not.
  • Co-simulation Platform Extension: The platform can be extended to support experiments with connected and automated vehicles (CAVs) by integrating autonomous driving stacks.
  • Countermeasure Design: Further research is needed to design more effective countermeasures for drivers who tend to ignore warnings, such as audio or vibration alerts, or even automatic emergency braking.
  • Real-world Validation: While the study used a real-world roundabout as a basis, the experiment was conducted in a simulator. Future work should aim for real-world validation.

Summary updated

BibTeX
@misc{zhang2023evaluation,
  author = {Cong Zhang and Chi Tian and Tianfang Han and Hang Li and Yiheng Feng and Yunfeng Chen and Robert W. Proctor and Jiansong Zhang},
  title = {{Evaluation of Infrastructure-based Warning System on Driving Behaviors-A Roundabout Study}},
  year = {2023},
  url = {https://openreview.net/forum?id=b0ihT7N96P},
}
Figure