Emotion recognition is attracting great interest for its potential application in a multitude of real-life situations. Much of the Computer Vision research in this field has focused on relating emotions to facial expressions, with investigations rarely including more than upper body. In this work, we propose a new scenario, for which emotional states are related to 3D dynamics of the whole body motion. To address the complexity of human body movement, we used covariance descriptors of the sequence of the 3D skeleton joints, and represented them in the non-linear Riemannian manifold of Symmetric Positive Definite matrices. In doing so, we exploited geodesic distances and geometric means on the manifold to perform emotion classification. Using sequences of spontaneous walking under the five primary emotional states, we report a method that succeeded in classifying the different emotions, with comparable performance to those observed in a human-based force-choice classification task.