[vorpalhex]

I think there are some valid concerns here. Someone running a clicker is already common procedure at some busy places, and it's often a hostess or someone else who is already "doing stuff".

Likewise any sensor is going to open up some questions about the data it collects and it's method of action - what's the error margin?

I think there's arguments for systems like this as well, but it's on you to make them, not just complain about others being skeptical.

[afar]

Accuracy is fundamental. When installed properly, a 99%+ f1 score is achievable.

[pbourke]

First off - this is really neat, especially the focus on privacy and the use of AI at the edge.

It seems like the occupancy limit problem for small rooms is much more challenging than the building-density problem, since a single error could DOS the room.

How do you account for that?

[afar]

Def. We call that the OB1 problem (off-by-one). If either event detection (did something happen) or event classification (entry or entrance) is wrong, you can be off the whole day. To solve that, we have a different approach / non-threshold approach for boolean occupancy. Not implemented here.

[vorpalhex]

What data does the sensor collect? How visible is the data it sends back to the platform?

[afar]

It's a custom sensor in the lidar family. Uses infrared lasers as illumination and generates depth data. Essentially millions of height values. When depth is rendered to be human legible it look looks like greyscale silhouettes (dark grey is far away, lighter gray is nearer). Sensor processes those values on the edge and published +1/-1 and telemetry data (system health).