DETECTION ZONE GRID 'mapping" Cam v4 -- Have you done any testing?

[carverofchoice]

Related to detection zone precision…I don’t believe they are precise in the way you are thinking. I know they definitely are not as far as the AI detections and notifications go, and I suspect it’s the same for motion event triggers too. There is a logical rationale behind why. I’ll give some illustrations to help show what I mean, but it should be mentioned that there is likely a little bit of variation between device models.

@carverofchoice

Let’s say that someone has a detection zone set up based on this random view that I pulled off the Wyze website (from a review):

• The green zone is the detection zone.
• Yellow is a person who is 30-50% in the zone and half or more outside the zone.
• Blue is a car driving by, completely outside the detection zone,
• Red is the camera’s estimate of the motion tag or parts that could possibly be related to the motion object (called the object “bounding box” outline…to get an idea of what I mean, it’s similar to what you see when you turn on the Green motion tag for the cameras to have it highlight an object in a green box…it’s not an exact fit/outline on the object, but a box shape that goes out a little farther).

The detection zone will trigger for the yellow guy and detect it as a person even though half the person is not in the detection zone, because the AI will consider a full object even if part of it is outside the detection zone, as long as some part of the object’s bounding box is within the detection zone.

Similarly, even though the blue highlight car is 100% outside of the detection zone area, the camera may consider the whole red highlighted area as part of the car and consider that within the detection zone area and the AI would also detect that car because part of the estimated bounding box was within the detection zone even though 100% of the car never was. So in this case, it can detect something fully outside of the detection zone in certain cases. To compensate for this, it is often suggested to give an extra buffer and shrink the detection area down an extra layer or 2 to help ensure undesired bounding box overlap into the detection zone is less likely.

In some cases, it is possible that the main motion is outside the detection zone, maybe even across the street, but then wind blows some foliage around and while it may not be as dramatic as the motion across the street (so the green motion tag stays mostly on the more dramatic motion elsewhere), it might be enough to trigger an invisible bounding box to overlap into the detection zone.

So they are precise in a way…as long as the object bounding box is fully outside the detection zone, then the object is not detected. BUT if any part of the object bounding box does overlap into the detection zone (even though we think the object itself never actually overlaps the detection zone), then it may be detected.
I know the AI has been confirmed to work this way by a Wyze employee, and I suspect any motion does as well. The REASON Wyze does it this way for AI objects is because if a person or other AI object is only HALF in a detection zone, then the AI wouldn’t recognize they are a person.

For example in the following pictures, if the AI could only analyze the legs of the person here (in the detection zone) but not the upper body/head (outside the detection zone), it wouldn’t realize it’s a person, so it has to be allowed to analyze things outside the detection zone IF part of the object is within the detection zone:

The problem is determining for sure what all is part of the object in question. This is a lot trickier than most people think, both with still frames and with inter-frame comparisons where some pixels of the object might change between 2 frames but not all of them at the same time depending on the speed and FPS, object size, color variation, etc. It’s a complex issue determining what all is part of an object or not in some cases.

So yes, Wyze could limit their motion detection to be strict lines, and I think they have started moving more in this direction for the initial motion event trigger (I think they should improve this)…but there are still some complicated considerations to take into account as well when it comes to AI detections.