Research on drowsiness detection in UAV operators based on the random decision forest method

Area of Science:

• Human-Computer Interaction
• Artificial Intelligence
• Cognitive Science

Background:

• Drowsiness is a major safety risk in critical operations, including unmanned aerial vehicle (UAV) piloting.
• Behavioral indicators like eye closure and head pose are used for drowsiness detection, but complex models lack interpretability.
• Understanding model decision-making is crucial for ensuring the reliability of automated safety systems.

Purpose of the Study:

• To utilize a Random Forest model as a diagnostic tool for analyzing dataset biases and feature correlations in drowsiness detection.
• To demonstrate how an interpretable framework can provide actionable insights into feature importance and model decision boundaries.
• To offer a method for auditing training data and improving the application of high-performance black-box systems.

Main Methods:

• Employed a Random Forest model not just for classification, but as an analytical tool.
• Utilized established benchmarks for evaluating model performance and interpretability.
• Analyzed feature importance and model decision boundaries to identify dataset characteristics and model behavior.

Main Results:

• The interpretable Random Forest framework successfully identified dataset biases and feature correlations.
• Actionable insights into feature importance and model decision boundaries were derived.
• The approach provided a method for auditing training data, highlighting potential issues.

Conclusions:

• Model interpretability is vital for developing trustworthy drowsiness detection systems.
• An interpretable framework can audit training data and inform the reliable use of complex models.
• This approach enhances the robustness and trustworthiness of drowsiness detection in operational environments, particularly for UAV operations.