Tracking Humans using Prior and Learned Representations of Shape and Appearance