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

Wojtowicz, Konrad; Wojciechowski, Przemysław; Panasiewicz, Adrian · 2026 · DOAJ

DOI: 10.1038/s41598-026-39195-y

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the critical safety issue of drowsiness in unmanned aerial vehicle (UAV) operators, proposing a detection system based on behavioral parameters and the Random Decision Forest method. The authors motivate their work by highlighting the severe risks of fatigue in aviation, citing historical accidents and the limitations of existing detection methods. While physiological methods offer high accuracy, they require invasive sensors, and vehicle-control-based methods only detect drowsiness after performance has already degraded. Consequently, the study focuses on non-invasive behavioral indicators, such as eye closure and head pose, to enable early detection. A key objective is to utilize the Random Forest model not merely as a classifier, but as an interpretable diagnostic tool to analyze feature importance and dataset biases, thereby providing transparency often lacking in complex black-box models. The methodology involves a real-time system that processes video feeds to extract specific facial features using the MediaPipe library, which detects 478 facial landmarks. The system calculates four primary parameters: Eye Aspect Ratio (EAR), Percentage of Eyelids Closed (PERCLOS), Mouth Aspect Ratio (MAR), and Euler head tilt angles (pitch and roll). Other potential indicators, including yawning, pupil diameter, and saccadic movements, were evaluated but excluded due to lack of predictive contribution, technical limitations in landmark detection, or signal noise. The extracted features are fed into a Random Forest classifier to determine the operator's drowsiness state. The system provides visual alerts via a graphical user interface and archives data for further analysis. The findings demonstrate that the selected behavioral parameters—specifically EAR, PERCLOS, MAR, and head tilt—serve as reliable indicators for drowsiness detection. The analysis revealed that yawning did not contribute to prediction accuracy, while pupil diameter and saccadic movements were deemed unreliable for this specific implementation. By employing the Random Forest model, the authors achieved a framework that allows for the auditing of training data and the assessment of feature correlations. This interpretability provides actionable insights into which features drive the model's decisions, contrasting with less transparent deep learning approaches like CNNs or LSTMs discussed in the literature review. The significance of this work lies in its contribution to developing robust, trustworthy drowsiness detection systems for operational environments. By prioritizing model transparency, the study offers a method to validate and understand detection algorithms, which is crucial for safety-critical applications like UAV operations. The approach underscores the value of interpretable machine learning in identifying dataset biases and ensuring reliable performance, ultimately supporting the creation of safer monitoring systems that can alert operators before fatigue leads to hazardous outcomes.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms