Spatio-Temporal Relationship Match: Video Structure Comparison for Recognition of Complex Human Activities