Analysis on Filter Feature Selection Methods for Driver Drowsiness Detection Using Facial Electromyography(EMG) Signals

Electromyography (EMG) signals are extensively investigated inputs for driver drowsiness detection algorithms due to its ability to monitor muscle activity. However, the use of facial muscle of EMG signals in combination with filter based feature selection methods has not been explored. This work examines several filter feature selection methods and its effects on the performance of EMG based drowsiness detection models, with a focus on frequency domain features. 12 samples were used which each sample will drive on driving simulator in a laboratory setting. BioRadio sensors were used to collect facial muscles data (orbicularis orris and masseter muscles). Three filter feature selection methods; Chi-square, MRMR, and F-test were employed to analyse seven frequency domain features of the EMG signals. These features were evaluated using classifiers such as kNN, LDA (pseudolinear), and SVM. The Chi-square filter, when paired with the Polynomial Support Vector Machine (SVM) classifier, produced the highest accuracy of 91.67% with eight features from both muscles. This displays the Chi-square filter's ability to identify relevant features with high accuracy. In conclusion, the orbicularis muscle and Chi-square filter was found to be the most effective model for feature selection in facial EMG based driver drowsiness detection.