Emotion Analysis through Circular Mapping of Postural Features Based on Russell’s Circumplex Model

This study proposes a regression-based emotion estimation method by mapping postural features onto Russell’s circumplex model using graph neural networks (GNNs). By extracting both geometric and structural features from skeletal coordinates and embedding them into a two-dimensional emotion space, we achieved visual and interpretable emotion representations. Numerical experiments on two public datasets (MASR and MEED) demonstrated that postural features, particularly those along the arousal axis, effectively reflect emotional states. Furthermore, mapping to a circular emotion space reveals that emotions like fear and sadness exhibit distinct postural patterns, whereas anger and happiness often overlap with surprise. These results suggest that our method enables a nuanced understanding of postural emotion expression and supports visual analysis of emotional similarities and distinctions.