Abstract
The visual system is able to extract from human movements subtle style information such as the expressed emotional state. There has been much interest in identifying the actual physical features supporting this perceptual ability. General level of movement activation is an important cue, but it does not support the distinction between, e.g., anger and happiness, two emotions sharing a similar level of activation. We motion-captured the gait of 25 individuals who walked neutrally at different velocities, and who expressed different affects (anger, happiness, fear, sadness). By fitting kinematic models the average flexion angles were computed as a measure of body posture. Dynamic information was extracted from the joint-angle trajectories using a blind source separation algorithm that results in highly compact representations of the trajectories (Omlor & Giese, NIPS 2006). Movements were presented by animating 3-D avatar stimuli, used in a classification and an expressiveness-rating experiment. The expressed emotion was recognised at rates between 70 and 90%. Anger and happiness, with high activation (fast and large movements), tended to be confused with each other, as did the less activated affects fear and sadness. These findings were in accordance with differences between these styles observed on the dynamic measures. In terms of body posture, arm flexion was a very expressive cue for both fear and anger, whereas happiness and sadness were associated with little mean limb flexion. Sadness and happiness could further be distinguished on the basis of head and spine inclination. Our findings show that the perception of emotional body expression is driven by quantifiable characteristics of the movement trajectories. Whereas activation is a basic feature distinguishing body expressions of happiness and anger from those of fear and sadness, postural tension could be an additional feature that is suitable for discriminating between emotions falling at the same end of the activation dimension.
Supported by HFSP, EC project COBOL, the Volkswagenstiftung and Hermann and Lilly Schilling Foundation.