Classification of Steering Wheel Contacts from Electrocardiogram Signals Using Machine Learning

Abstract

With the current day advancements in both computational power and machine learning (ML) techniques, there is a fundamental shift toward the application of new and smarter technologies. Worldwide incidents of motor vehicle crashes cause financial and emotional distress, along with physical injury, and even death, often stemming from driver fatigue. Nowadays, advanced ML techniques can be combined with electrocardiogram signals recorded from hand-contact with the motor vehicle steering wheel, to accurately detect the onset of driver fatigue. However, the signal recorded is only viable for fatigue analysis when two hands are in contact with the wheel. This work aims to carry out a comparative evaluation on a selected set of ML algorithms, when considering their ability to determine the number of contacts on a steering wheel. The ML classifiers considered in this study include the unsupervised methods K-means clustering, and Gaussian Mixture Model, and the supervised methods Support Vector Machine (SVM), Linear Discriminant Analysis, and Convolutional Neural Network (CNN). The evaluation is carried out based on both standard ML evaluation metrics including accuracy, precision, specificity, and computational cost. The experimental results show that the CNN produced the highest-level accuracy (>99%), but also had the highest computational cost. The SVM method presented the most balanced performance with a low computational cost and the second highest-level accuracy (94%). This paper assesses the viability of ML algorithms to eliminate the non-viable segments within ECGs that are used to determine driver fatigue. This is done by evaluating the techniques ability to consistently detect the number of contacts on a steering wheel, and its ability to be implemented in real-time, through the analysis of computational cost.