Leveraging driver vehicle and environment interaction: Machine learning using driver monitoring cameras to detect drunk driving

Document Title

Leveraging Driver Vehicle and Environment Interaction: Machine Learning Using Driver Monitoring Cameras to Detect Drunk Driving

Document Information

• Subject Area: Human-Computer Interaction, Machine Learning, Intelligent Driving
• Keywords: Drunk Driving Detection, Health and Safety, Driving Behavior, Machine Learning, Driver Monitoring Cameras, Eye Movement, Head Movement

Research Background and Problem Statement

• Problem and Challenges: Globally, alcohol-related driving issues contribute significantly to health damages and road traffic accidents. Traditional drunk driving detection methods, such as ignition interlock devices, are costly and complex to maintain, limiting widespread adoption. Existing technologies for detecting driving behavior struggle to account for individual differences and variations in driving scenarios.
• Significance: Alcohol's impact on driving behavior is a critical issue in traffic safety. Reducing drunk driving incidents to mitigate threats to life and promote healthy behavior changes holds significant societal importance.
• Research Motivation and Related Work: The authors observed that driver monitoring cameras are increasingly becoming standard equipment in modern vehicles, prompting an investigation into their potential for detecting drunk driving. Existing studies primarily focus on driving data (e.g., steering wheel operations) for drunk driving detection but lack utilization of camera data.

Solution

• Method or Solution: A machine learning system based on driver monitoring cameras is proposed, which predicts the extent of alcohol influence by analyzing the driver's eye movement and head activity.
• Innovations:
  • The first use of driver monitoring cameras for drunk driving detection, addressing the accuracy limitations of traditional driving data-based methods.
  • The system demonstrates strong generalization capabilities across unseen individuals and driving scenarios.
  • The model leverages known pathophysiological effects of alcohol on visual and motor functions for validation, enhancing interpretability.
• Implementation Steps:
  1. Real-time capture of the driver's eye movement and head movement data using driver monitoring cameras.
  2. Feature extraction via a sliding window method to generate input datasets for the machine learning model.
  3. Application of classification tasks to train the model to predict two blood alcohol concentration thresholds: a low concentration level for early warning and a high concentration level exceeding legal limits.

Research Outcomes

• Specific Results:
  • The system successfully detected alcohol influence in driving simulator experiments and provided warnings for two alcohol concentration levels (early warning AUROC: 0.88, high concentration detection AUROC: 0.79).
  • Experiments confirmed that camera data offers higher predictive accuracy compared to traditional driving behavior data, with camera data alone outperforming its combination with vehicle CAN signal data.
• Advantages:
  • Low-cost system, easy to implement on a large scale.
  • Driver monitoring cameras are increasingly standard in modern vehicles, aligning with future regulatory requirements.
  • Provides real-time feedback for early intervention in drunk driving.
• Experimental or Evaluation Results:
  • The system performed consistently across various driving scenarios (highways, urban roads, and rural roads) in the simulator.
  • Demonstrated robust generalization across individuals and scenarios during evaluation.
• Limitations and Future Directions:
  • The system has not yet been tested in real driving environments and requires further calibration to reduce false positive rates.
  • The model may be influenced by individual differences, such as drinking habits or visual behavior issues.
  • Future research is recommended to include real-world road experiments and validation across broader target groups (e.g., different ages or health conditions).

Data Availability

Requests for original data must undergo review by the Scientific Research Ethics Committee and will only be accepted for non-commercial purposes. All shared data will be de-identified and require signing a material transfer agreement.

Acknowledgments

The research was funded by the Bosch IoT Lab, jointly organized by the University of St. Gallen and ETH Zurich, and supported by the Swiss National Science Foundation. The funding bodies did not influence the study design, implementation, or data analysis.