Skip to content

Bug (ish) Report : AI motion detection "sees" things that are not in motion

edited April 29 in SecuritySpy

OK - so I get this one - and I know just enough about AI / machine learning models to understand what's happening... so it really depends on what the API allows you to do or not do... but here's the scenario.

Camera in garage - cars parked not moving at all.

Person walks into garage - that triggers motion - AI determine it's a human... AND a vehicle. Vehicle was there before event and after event - never moved - but the AI doesn't understand that as it analyzes "a frame with motion" and determines it's got a human and a vehicle in it - so it's gotta be both.

So "bug" - because the vehicle didn't move... but also "ish" because, well, machine learning...



Comments

  • Your summary is basically correct. The current neural network that we have developed for SecuritySpy is a classifier that works on still images. It is extremely accurate at detecting whether there is a human or a vehicle present in the image (96% and 97% accuracy respectively), but it can't tell whether that human/vehicle is moving. Therefore, if you have a situation as you describe, where there is a stationary vehicle, and something else moves in front of the vehicle, this will trigger SecuritySpy's vehicle detector.

    I wouldn't call this a bug, but rather a known limitation of the type of neural net SecuritySpy uses. In the future it may be possible for us to develop an AI that takes a sequence of images, which it can then process to detect only moving objects. But this is far more difficult to develop, requires a whole new set of training data, and would run far slower than the current network.

    What this means in practice is that you should not use the vehicle detector for a camera whose view normally includes stationary vehicles.

  • Yeah. I figured. Computers are not smart ;)

  • Yes. At least, computer's aren't smart in the way that humans are smart :)

Sign In or Register to comment.