Very interesting! I could see this being very difficult to implement. For example, if I have a detection zone that covers just my sidewalk (perpendicular/horizontal to the camera pointing direction), then in some cases, only a person’s feet would technically be in the detection zone. Their knees and body and head/face would be outside the detection zone. Therefore, if a person walks on the sidewalk, it seems the AI would be told to analyze only the feet, and determine this is not a person because the body and head were not in the detection zone (which the AI was told to ignore), even though the trigger (a person’s foot) was in the detection zone.
Is this potentially part of what is causing the problem with limiting the AI to detection zone analysis only?
Example screenshots of my above description using detection zone grid:
Feet would be in the detection zone, triggering an event, but the body and head would be excluded from AI detection, thus causing the AI to not recognize a “person”:
AI here would only recognize a “PET” because my body/head are excluded from the detection zone:
(The AI would still recognize my “pet” in theory…though in my case, it would probably think my black cat is a “person” but that’s beside the point here 
)
If the AI is told to ignore everything outside the detection zone, then anyone could walk right across my front sidewalk without me ever getting a person detection because the AI would then be told to ignore the body/head which it has been trained to look for.
I guess you could try to tell the AI to check anything connected to the triggering object (my legs) and see if there is a body/head connected to that triggering object that started in the detection zone, but when it comes to a picture, isn’t EVERYTHING connected to the triggering object in some way? I guess a person on the other side of the street would have separate legs/body/head not linked to the legs in the triggering object though, so the AI could be trained to ignore the non-connected person(s).
Anyway, very interesting. I would certainly love to hear anything you’re willing share related to the issue of AI analyzing inside/outside of detection zone limits to ensure the above examples aren’t problems. I can see you have your work cut out for you though! It sounds fairly complex to train the AI this way.