A Robust and Efficient Video Representation for Action Recognition