Exploring Contextual Information in a Layered Framework for Group Action Recognition