Physiological-based group emotion recognition : novel methods and real-world applications

Emotions determine human thinking and behaviour. Thus, affective computing is extremely relevant in several applications, from mental health and the creation of personalized services to entertainment. This work begins by exploring the state of the art in the area of intra-personal emotion recognition through physiological signals, making a quantitative analysis of the various existent approaches. Next, an interpersonal model was explored that incorporates group dynamics using the WGS methodology, where the synchrony between the physiological signals of a group was studied. This model presented superior results to existing methods and identified relevant synchrony measures. However, the generalization of group analysis was limited by a lack of databases. To overcome this limitation, a hardware and software infrastructure was developed for the acquisition of physiological signals in groups — the EmotiphAI platform, which includes wearable devices, a collection centre, and a user interface, enabling the simultaneous collection of data from 10 devices at 60 Hz.

Subsequently, a retrospective annotation system was incorporated that selects relevant physiological segments, which demonstrated high accuracy and user satisfaction. This functionality paves the way for the annotation of longer-duration content for the creation of naturalistic datasets. Using EmotiphAI, the G-REx dataset was created with annotated physiological signals from 190 subjects across 31 movie sessions. Moreover, the respective protocol validated the use of EmotiphAI in a real-world environment.

Overall, this thesis contributes to the advancement of applications in the area of affective computing in the real world through the development of: interpersonal algorithms for emotion recognition; tools for group physiological data acquisition and annotation; and a naturalistic dataset.