Driver fatigue detection using PPG signal, facial features, head postures with an LSTM model

Area of Science:

• Road safety
• Artificial Intelligence
• Biomedical Engineering

Background:

• Driver fatigue is a significant cause of road accidents.
• Existing detection methods using single features have limitations in accuracy and prediction efficiency.
• A novel, multi-feature approach is needed for robust fatigue detection.

Purpose of the Study:

• To introduce a novel driver fatigue detection framework integrating facial features, head pose, and photoplethysmography (PPG) signals.
• To develop and optimize a Long Short-Term Memory (LSTM) network for enhanced fatigue detection accuracy.
• To validate the proposed model through a real-world driving experiment.

Main Methods:

• Collected multi-source feature data from 30 drivers during a real-road driving experiment.
• Extracted 2D facial landmarks, 3D head pose parameters, and five-dimensional heart rate variability (HRV) features from PPG signals.
• Fused ten-dimensional features to create a dataset and trained/optimized an LSTM model using Momentum, Rmsprop, Adam, and SGD algorithms, comparing it with DT, RF, and BiLSTM.

Main Results:

• The Adam-optimized LSTM model demonstrated superior performance with 97.36% accuracy, 97.4% precision, and 97.4% recall.
• The proposed model significantly outperforms traditional methods and other evaluated algorithms.
• The model's effectiveness indicates its capability for timely and accurate fatigue prediction.

Conclusions:

• The integrated multi-feature approach using an LSTM network provides a highly accurate and reliable method for driver fatigue detection.
• The developed framework offers a promising solution for improving road safety by enabling early warnings for fatigued drivers.
• Further research can explore real-time implementation and integration into vehicle systems.