Machine Cognition of Violence in Videos using Novel Outlier-Resistant VLAD