AI Human Detection Behaviour Question
  • Hi -

    I'm on 5.0.2b6 and I have a question about motion detection, and specifically about the "AI: trigger Motion Capture at these detection thresholds" option. I have it set to 60 for humans.

    I also have "enable motion detection in video" set to 90, with a trigger threshold of one second.

    I was expecting that foxes playing in the garden would trigger because they are moving for more than a second - but also that events including people would be recorded due to the AI component. But since I've had the settings like this, I have missed the foxes altogether, even when I've been watching them with my own eyes through the window(!)

    So does the AI: Human.... setting preclude detecting everything that's not human? How can I detect the foxes and retain the ability to sensitively detect people in the back garden? Is this a case where really I should switch off the AI and get back to the days of swaying branches being indistinguishable from burglars and foxes?

    Some commentary on the 'cascade' of detections or the mutual exclusivity of the options would help me to optimise this, thanks.
  • The whole point of the human detection AI is to trigger on humans only, and ignore all other motion. So it sounds like the AI is working exactly as intended: it is ignoring all non-human motion, such as swaying branches and foxes.

    Currently, if you want to detect foxes, then you will have to turn off the AI. SecuritySpy includes a human detector and a vehicle detector, as these are the objects that will be of most interest in virtually all CCTV scenarios. Many users have also asked for an animal/pet detector, and we do intend to offer this in the future, when we have gathered enough training data to construct an accurate enough AI for this.
  • Thanks Ben - perfectly clear and common sense. :-)
  • Ben, I've noticed that the AI pick up on shadows of people walking past my driveway, the actual footpath is masked out, but their shadow crosses the detection area.

    Is this something that can be tweaked?
  • I've also had problems with shadows, in the late afternoon.

    I set the green mask so it covers everything from the sidewalk to the street. But SS still triggers, even with the human threshold set to 95, sensitivity to 50, and trigger time to 1 second.

    The person is masked out, but the moving shadow cast on the protected area will reliably trigger. I'm guessing the shadow moving at a velocity similar to a human is detected as a human.

    Looking at SS AI Prediction image files, the top of the red box extends far into the green mask area.

    Are there better settings I should be using?

  • Shadows are unfortunately difficult to deal with, for two reason. One is that a shadow of a person has many of the same features as a person, so to train an AI to tell the difference is difficult and may have undesirable outcomes (e.g. more false negative detections of real humans).

    The other reason is that the mask determines where the pixel-based motion detection looks, but after motion is detected by this, and the image is passed to the AI, the whole area of motion is used for classification, whether it is in the masked area or not. We designed it like this in order to reduce the possibility of false negative events: i.e. if a person is moving, and is half covered by the mask, if were were to run the AI on just the unmasked area it may not be able to classify this partial image as human. But by running the AI on the whole area of movement, it is much more likely that the human gets correctly classified and a trigger generated.

    There is an inevitable tradeoff between false negatives (no trigger generated when you want one) and false positives (a trigger generated when you don't want one). Since false negatives are much worse than false positives for CCTV applications, we have set up the system to minimise false negatives. Unfortunately this does mean that certain events will cause false positive detections.
  • Thanks for the heads-up, Ben!

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!