Anomaly detection applied to video analysis

I was wondering how can we apply Anomaly detection (AD) to video analysis, it is difficult for me because almost all algorithms for video are supervised, and they can’t go beyond their training knowledge, in that sense detecting anomalies is very hard in this scenario.
for example, imagine a crowd of people in some place, metro, bus, bank, etc, could we detect a person with a different behavior? is there another o similar algorithm to AD for this?
maybe for some scenarios, we could apply position tracking, and from this apply AD.


The lecture’s approach requires us to have quantiative features, then we model each feature with a distribution, then we spot which sample has features that are at the tails of those distributions.

Now, how are you going to quantify “behavior” in useful ways? I think this is what you might need to think about.


Seems to me that you’d first have to identify the objects in a video frame, and then (as Raymond suggests) invent some method to measure anomalous behavior.

Object detection is not simple, but it’s covered in some depth in the Deep Learning Specialization.


What you suggest is a case of easier said than done.

If you want to apply the technique of anomaly detection rather than a supervised learning technique, it would require you to have the kinds of behaviour that are highly probable and those that are highly improbable. Now to make matters more complicated, these behaviours will need to be identified from a video feed. So, this again circles back to the need for image recognition which is a supervised technique.