An XGBoost approach to detect driver visual distraction based on vehicle dynamics

Area of Science:

• Road Safety
• Machine Learning Applications
• Automotive Engineering

Background:

• Distracted driving, particularly involving mobile phone use, significantly elevates the risk of severe road accidents.
• Existing methods for detecting driver distraction often require additional sensors or are limited in accuracy.

Purpose of the Study:

• To develop and evaluate an XGBoost model for detecting visual distractions using only vehicle dynamics data.
• To identify optimal parameters for the distraction detection model, including time window size and feature selection.

Main Methods:

• A simulated driving experiment was conducted with 36 participants.
• Vehicle dynamics data were processed using time window and Fast Fourier Transform methods, yielding 26 parameters.
• The performance of the XGBoost model was optimized by varying time window sizes (1-7s) and the number of input features.

Main Results:

• The optimal time window was determined to be 5 seconds with 23 input indicators.
• The XGBoost model achieved high performance metrics: 85.68% accuracy, 85.83% precision, 83.85% recall, 84.82% F1 score, and 0.9319 AUC.
• Key features for distraction detection included the standard deviation of vehicle sideslip rate and steering wheel angle spectrum components.

Conclusions:

• Vehicle dynamics, specifically steering wheel angle and sideslip angle, are strong indicators of driver distraction.
• The proposed XGBoost model shows promise for integration into advanced driving assistant systems (ADAS) to mitigate risks associated with distracted driving.