Detecting abnormality with separated foreground and background: Mutual Generative Adversarial Networks for video abnormal event detection