Automatic emotion recognition remains a central challenge in affective comput- ing due to the subjective nature of emotional experience, high inter-individual variability, and the limitations of traditional self-reporting methods. Existing approaches are often constrained by discrete emotion representations and by a limited consideration of individual differences. In contrast, emotional states can be assessed continuously according to Russell’s circumplex model, which represents affect along the dimensions of arousal and valence rather than discrete categories. Building upon this continuous formulation, this thesis proposes a multivariate fusion framework for affect estimation in the Valence–Arousal space, integrating physiological signals, namely Galvanic Skin Response (GSR) and Heart Rate (HR), Pupillometry, and Facial Emotion Recognition (FER) within a unified modeling pipeline.