Abstract

Movement without motion - the role of body postures in action recognition and understanding.

The posture of a body is one of the key features for the analysis and recogntition of human actions.
Articulated (or biological) motion could be considered as a sequence of static snapshots containing expressive body postures and their transitions into each other. These snapshots on their own carry enough information to enable the recognition of a human action. Moreover, isolated static snapshots of articulated body poses are sufficient to evoke the percept of motion and even allow an anticipation of the associated action. Evidence suggests that the representation of animate objects is built in cortical region STS. It is still an open issue to which extent form and motion information contribute to the generation of such representations. We aim at proposing a neural model to learn sequence and motion patterns, explaining neural activations of implied motion. The proposed model consists of two mainly dissociated processing streams, segregated into the ventral and dorsal pathway. Along the ventral pathway, local form features are used to build complex form patterns. Likewise, in the dorsal pathway, local optical flow features are utilized to generate short-term optical flow patterns. Both pathways converge in model area STS, where the interaction of shape and motion representations is enabled via lateral interconnections. Sequence selectivity is achieved by integrating the activities of STS cells over time. The learning throughout the model is realized unsupervised via Hebbian plasticity. The model cell representations have successfully learned to recognize (human) action sequences. Static images, that appear at the beginning or end of a trained sequence evoke higher activities in model STS cells and are interpreted as anaolgue representations of intentions. Next to the presentation of the proposed neural model for the recogntition of implied motion, I will give a short overview of the relevance of body postures for action recognition and, as well, social signal processing in general and show some preliminary results on the detection and interpretation of stick-figure-like representations of a human bodies.