Using long short term memory and convolutional neural networks for driver drowsiness detection.

Area of Science:

• Computer Science
• Artificial Intelligence
• Transportation Safety

Background:

• Driver fatigue is a significant cause of road accidents, necessitating effective detection methods.
• Existing drowsiness detection methods like electroencephalogram (EEG) are accurate but intrusive, while vehicle dynamics are less accurate.
• Eye movement analysis offers a balance but typically requires expensive eye-tracking systems, hindering practical implementation.

Purpose of the Study:

• To develop a practical and accurate driver drowsiness detection system using readily available eye images.
• To leverage Recurrent Neural Networks (RNNs), particularly Long Short-Term Memory (LSTM) models, for analyzing eye movements.
• To compare the performance of 1-D LSTM (R-LSTM) and Convolutional LSTM (C-LSTM) for drowsiness detection.

Main Methods:

• Utilized eye image patches from 38 subjects in a simulated driving experiment.
• Employed 1-D LSTM (R-LSTM) as a baseline and Convolutional LSTM (C-LSTM) for direct 2-D image analysis.
• Validated drowsiness levels using simultaneous electroencephalogram (EEG) power spectral analysis to generate ground truth labels.

Main Results:

• The R-LSTM approach achieved an accuracy of approximately 82%.
• The C-LSTM approach demonstrated superior performance, with accuracy ranging from 95% to 97%.
• The proposed LSTM-based methods significantly outperformed a recent eye-tracking based approach.

Conclusions:

• Directly analyzing eye images with LSTM models provides a highly effective and practical solution for driver drowsiness detection.
• Convolutional LSTM (C-LSTM) shows particular promise for real-time applications due to its high accuracy and ability to process 2D images.
• This approach offers a cost-effective alternative to traditional eye-tracking systems, enhancing road safety.