Stress is a factor that affects many people today and is responsible for many of the causes of poor quality of life. For this reason, it is necessary to be able to determine whether a person is stressed or not. Therefore, it is necessary to develop tools that are non-invasive, innocuous, and easy to use. This paper describes a methodology for classifying stress in humans by automatically detecting facial regions of interest in thermal images using machine learning during a short Trier Social Stress Test. Five regions of interest, namely the nose, right cheek, left cheek, forehead, and chin, are automatically detected. The temperature of each of these regions is then extracted and used as input to a classifier, specifically a Support Vector Machine, which outputs three states: baseline, stressed, and relaxed. The proposal was developed and tested on thermal images of 25 participants who were subjected to a stress-inducing protocol followed by relaxation techniques. After testing the developed methodology, an accuracy of 95.4% and an error rate of 4.5% were obtained. The methodology proposed in this study allows the automatic classification of a person's stress state based on a thermal image of the face. This represents an innovative tool applicable to specialists. Furthermore, due to its robustness, it is also suitable for online applications.
Keywords: face detection; face landmarks; machine learning; short TSST; stress; thermography.