Design of Early Fatigue Detection Elements of a Wearable Computing System for the Prevention of Road Accidents

Abstract

This paper presents the summary of research involving two state Neural Network classifiers [4] specifically designed for the detection of fatigued state of a driver. Several physiological parameters such as Skin Conductance (SC) and Oximetry Pulse (OP) were considered individually as well as in combination, to design the fatigue classifiers using Multilayer Perceptron Neural Networks (MLP NN) with one and two hidden layers, and the Support Vector Machine (SVM). Performance analysis of the classifiers has been carried out using independent validation method and the Receiver Operating Characteristic (ROC) method. Performance indicators used were Percentage Classification Accuracy (PCLA), Mean Square Error (MSE), Normalized MSE (NMSE), Area under ROC curve (AROC), Area under Convex Hull of ROC Curve (AHROC), Sensitivity (S), Specificity (R), Predictive Pre and Predictive Post. From the comparative analysis of the classifiers, it is evident that the two hidden layer MLP NN gives the best classification accuracy at hidden layer comprising of 65 and 80 Processing Elements (PE) respectively when the combined feature matrix of SC and OP was used as input to the network.