A classical approach to abnormal activity detection is to learn a
representation for normal activities from the training data and then use this
learned representation to detect abnormal activities while testing. Typically,
the methods based on this approach operate at a fixed timescale