Moving Beyond the Simulator: Interaction-Based Drunk Driving Detection in a Real Vehicle Using Driver Monitoring Cameras and Real-Time Vehicle Data

Research Background and Problem

• What problems or challenges did the authors identify?
  Alcohol consumption has become a serious public health and traffic safety challenge. Globally, more than 700 people die daily in traffic accidents caused by drunk driving, accounting for 22% of all traffic-related fatalities. However, existing alcohol detection methods (e.g., in-vehicle ignition interlock breathalyzers) are costly and do not meet the "passive" requirements of future autonomous driving technologies. Additionally, the applicability of existing drunk driving detection research based on simulated test environments has not been fully validated in real vehicle settings.

• Why is this problem important?
  Drunk driving poses risks not only to individual health but also to traffic safety. Developing scalable and cost-effective solutions for detecting drunk driving can significantly reduce related accidents and accelerate the full deployment of autonomous driving technologies.

• Research Motivation and Related Work
  Although some studies have attempted to develop machine learning (ML) models to detect drunk driving behaviors using simulator data, the real-world feasibility and generalizability of these results remain contentious. Active safety systems, such as driver monitoring cameras (DMC) and vehicle communication networks (CAN), provide a technical foundation for real-time drunk driving detection. However, research on detecting drunk driving in real vehicles and without observational samples has been scarce.

Solution

• What methods or solutions did the authors propose?
  The authors developed and evaluated a machine learning system that combines driver monitoring cameras (DMC) and real-time vehicle communication data (CAN) to detect drunk driving behaviors. This system extends prior simulator-based work and leverages standard automotive hardware.

• What are the innovative aspects of this solution?

  1. Real-world applicability of the model: For the first time, a drunk driving detection system was validated in real vehicles rather than in simulator environments.
  2. Data fusion: Integration of DMC and CAN data to enhance the model's ability to detect intoxicated states.
  3. Introduction of new features: Development of new features based on eye movement region events, further improving the model's adaptability to real driving environments.
  4. Rigorous experimental design: Inclusion of placebo and reference groups to eliminate interference from non-target factors (e.g., training effects and fatigue).

• What are the implementation steps and key technologies used?

  1. Data Collection: Conducted randomized controlled trials on a test track in Switzerland, recruiting 54 participants and recording their driving data under varying blood alcohol concentrations.
  2. Feature Extraction: Extracted 580 statistical features from DMC and CAN data, including head movement, eye motion, electronic control interactions, and vehicle dynamics features.
  3. Modeling: Used sliding windows and logistic regression (L1 regularization) for classification tasks, including "drinking or not" and "exceeding the World Health Organization (WHO) recommended blood alcohol concentration (BAC) limit of 0.05 g/dL or not."
  4. Model Validation: Assessed the model's generalization to unseen driver data using leave-one-subject-out cross-validation.

Research Findings

• What specific results were achieved?

  1. The model combining DMC and CAN data achieved an AUROC (Area Under the Receiver Operating Characteristic curve) of 0.84 ± 0.11 for the "drinking or not" task and 0.80 ± 0.10 for the "exceeding BAC limit" task.
  2. The model's performance remained stable after introducing control groups, demonstrating its resistance to potential biases.
  3. The approach captured known physiological patterns of drunk driving, including changes in eye movement and vehicle handling behaviors.

• What advantages does it have compared to existing solutions?

  1. The system's performance in real vehicle environments was comparable to simulator studies, validating the transferability of simulation data.
  2. It does not require high-precision laboratory equipment and can detect drunk driving using only standard industrial-grade DMC and CAN hardware.
  3. The method is cost-effective, as all processing can be completed within an in-vehicle closed-loop system without requiring additional data storage or transmission.

• What were the experimental or evaluation results?

  • Performance was lower when using DMC or CAN data alone, but classification ability significantly improved when the two were combined.
  • The model exhibited stable performance across different driving scenarios (highways, rural roads, urban areas).

• Limitations and Future Directions

  1. Data Diversity: Although the study sample was balanced in terms of gender and age, it did not include sufficient racial diversity, which may affect the generalizability of eye movement features.
  2. Environmental Realism: The test track lacked the complex and variable driving conditions of the real world (e.g., nighttime driving, interference from other road users). Future work should use real-world road data to further validate the model's robustness and applicability.
  3. Impact of Autonomous Driving: As partial autonomous driving features are introduced, the relevance of CAN data may decrease. The model needs to be enhanced and validated for highly automated driving modes.

Conclusion

This study significantly advances the transition of drunk driving detection from simulators to real-world applications by incorporating real-time vehicle data and standard automotive hardware. The model not only demonstrated robust performance in experiments but also paved the way for active safety systems in future autonomous vehicles.