Fundamental issues on the recognition of autonomic patterns produced by visual stimuli

One of the common approaches to the automatic emotion recognition problem is based on biological signal analysis. In this context, this paper aims at identifying the biological component related to levels of arousal of the subjects, and to use such a component to automatically discriminate among these levels. We have formalized the automatic emotion recognition as a classification problem. In order to allow the system to generalize over different subjects, we addressed two crucial aspects of the procedure: normalization and cross-validation. Assuming that different subjects could react differently to the same stimuli, we defined a distance metric between their models. We performed an experiment where 14 volunteers were stimulated by means of the IAPS set of pictures divided into classes of increasing arousal intensity. Under the effect of these external visual stimuli, subjects exhibited a principal component in their autonomic space that accounts for full class separability. Moreover we observed that some of the subjects' data can be represented by the same model, while others have to be represented differently possibly due to a poor induction mechanism. This work demonstrates the possibility to build a model able to generalize over different subjects without over-fitting, but we have to guarantee that data used to build the model represent sufficiently well the measured phenomena.