Deep Learning-Based Classification of Stress Levels during Real-World Driving Tasks

Abstract

Stress has been identified as one of the contributing reasons to vehicle crashes, which cost governments and society a large amount of money in terms of lost lives and productivity. Any alteration that creates physical, emotional, or physiological strain when driving is referred to as driving stress. Driving stress may vary depending on the different road conditions of driving. Understanding drivers’ discontent is one of the most important areas for improving intelligent transportation systems over the existing system. This study presents methods for analyzing and classifying EMG data collected during real-world driving tasks at different driving locations using a convolutional neural network (CNN). In this paper, there are 9 subjects (driver records) of at least 60 minutes duration. Developing CNN from scratch is difficult and it also demands specialized knowledge. As it was previously trained on the ImageNet dataset and could operate effectively with the small amount of training set, pre-trained CNN minimizes the effort of developing models from scratch. CNN is employed in the proposed work to classify driving stress levels by evaluating discriminatory patterns in spectrogram images. In the proposed work, the performance of pre-trained CNN SqueezeNet, GoogLeNet, and ResNet50 in identifying the level of stress (low, medium, and high) is compared. GoogLeNet performed best, with an accuracy of training and validation of 85.71% and 66.67%. Followed by ResNet50 with an accuracy of 71.43% and 66.67% and SqueezeNet with an accuracy of 71.43% and 55.56%, respectively.