Driving stress detection using physiological data with machine learning

Abstract

Stress is a problem that affects both physical and mental health, causing negative emotional states.  Stress can impair the driver’s ability to perceive and handle situations in driving safety. Therefore, the detection and assessment of stress levels play an important role in improving comfort, well-being, and enhancing the driving experience for drivers. Using the AffectiveROAD dataset, this paper proposes a method of classifying stress levels through physiological signals obtained from driving sessions. These signals are time-aligned and pre-processed to extract the suitable features within a five-second period. Based on the obtained features, Machine Learning models are trained to classify stress status into five levels. The tested results show that the accuracy reaches 94% with the Random Forests (RF) when using the seven most important features from the HR, EDA, TEMP signals, and 99% when incorporating the overlapping technique for 10-fold cross-validation.