Ability to run classification models, for example to recognize state if gate is open/closed

[michalsrb]

Describe what you are trying to accomplish and why in non technical terms
I understand this may be out of scope for Frigate, but it would be useful feature for me.

I would like to be able to detect whether my gate is open or closed. It is always present in the same place in the same camera, so I am not looking to detect whether a gate is there, but what is its current state. I expect different people's gates look wildly differently, so this would be best a custom model for each of such "rectangle".

Other use cases could be detecting things like if a flag is up or down, if bird feeder is full or empty, if a sandbox is covered or uncovered...

Describe the solution you'd like

• Ability to define rectangle on a camera and a set of custom states (in my case "open", "closed").
• Then to classify snapshots of that rectangle as one of the states or as "unknown" if it is for example obstructed.
• Then to train some kind of image classification model on it.
• Then when there is movement overlapping the rectangle, classify the image in the rectangle and expose the results over MQTT (ideally current state, last known state, time of last detection).

Describe alternatives you've considered
Alternative would be to pull the camera image to some other software and do the detection there.