Algorithms for Emotion Recognition

Technological advancements have increasingly facilitated emotion recognition through physiological sensors integrated into intelligent devices, such as earables and wristbands. People commonly wear these devices in their everyday lives (i.e., in the wild). Patterns can be extracted from various physiological signals, enabling the recognition of emotions. This capability can be integrated into diverse applications, such as attention management systems, human-robot interaction, and stress detection, enhancing them to support humans with an empathic component. This thesis addresses the algorithmic design, adaptation, and utilization of emotion recognition with physiological signals for the mentioned applications. Three key research gaps are identified and addressed: 1) the acquisition of accurate physiological data for emotion recognition in the wild, 2) the improvement of the emotion recognition process, and 3) the need for ethical standards that accompany performing emotion recognition using physiological sensors in the wild. By addressing the improvement of the emotion recognition process, this thesis focuses on mitigating the impact of physical activity on physiological sensor data measured outside the laboratory. Classification algorithms are trained on new mathematical features with data affected by physical activity to create enhanced emotion recognition models, achieving a classification accuracy of up to 73%. Finally, it is demonstrated that emotion recognition can enrich applications like attention management systems by finding opportune moments for interruptions based on smartphone notification response time prediction.