Extending conditional random field models for human action recognition