A self-organizing neural model for context-based action recognition

An artificial neural network model for visual recognition of actions is proposed. In contrast with the existing gesture recognition systems, our model learns to recognize "true" actions, i.e. object-directed actions with causal chains of events, such as "He threw the ball at the window and broke it". The core mechanism of our model consists of a triplet of parallel self-organizing networks; The first pair of networks learn and recognize "spatial relationship" and "movement patterns", whose output is integrated by the third "temporal context" network.