Detect car boot open/closed attribute when car is parked

[AndyFug]

Hi,

I am trying to use object classifications with attributes to detect when either of our car boots (trunks) have been left open. I don't think I can use object states for this because the cars can change position on our driveway etc. Ultimately I want to setup a notification in HA to say "The boot has been left open on the Mazda".

What I'm finding is that once the car has parked on the driveway and I open the boot, I don't think the object is being re-classified. I've tried leaving it for a period of time (5 mins or so), but it still doesn't seem to pick it up.

My stationary config is as per Frigate defaults. Cars are being detected OK on the front_driveway zone and I am just starting to train the classification for Family Cars, which seems to be going ok. It feels like the issue is that once the object becomes stationary, it is not being re-evaluated, but I maybe wrong.

Some relevant parts from the config:

cameras:
  driveway:
    review:
      alerts:
        labels:
          - person
          - car
        required_zones: front_driveway
    zones:
      front_driveway:
        coordinates: 
          0.082,0.492,0.434,0.273,0.791,0.669,0.993,0.895,0.953,1,0.213,1
        loitering_time: 3
        friendly_name: Front Driveway
        inertia: 5

classification:
    Family Cars:
      enabled: true
      name: Family Cars
      threshold: 0.8
      object_config:
        objects:
          - car
        classification_type: sub_label
    Car Boot Open:
      enabled: true
      name: Car Boot Open
      threshold: 0.8
      object_config:
        objects:
          - car
        classification_type: attribute

Any general advice would be greatly appreciated... I may be approaching this in the wrong way.

Cheers,

Andy

[inkeep-ai-assistant[bot]]

Your understanding is correct. Once an object becomes stationary, object detection does not continually run on that object. Detection will only run again when motion is detected near the object, or at the configured interval. (1)

The default stationary configuration is:

detect:
  stationary:
    interval: 50
    threshold: 50

(2)

The interval setting controls the frequency for running detection on stationary objects. By default, detection runs every 50th frame on stationary objects. (1)

To have Frigate re-evaluate your parked cars more frequently for attribute changes (like an open boot), you can decrease the interval value. For example:

detect:
  stationary:
    interval: 10
    threshold: 50

(2)

This would run detection every 10th frame on stationary objects instead of every 50th, giving your attribute classifier more opportunities to detect when the boot opens.

Keep in mind that lowering the interval will increase resource usage since detection runs more frequently on stationary objects. You may need to experiment to find a balance between responsiveness and system load.

📚 Sources:

1. Stationary Objects
2. Full Reference Config

---

Have another question? Just tag @inkeep.

[AndyFug]

Increasing the stationary.interval value does not seem to work. I just don't think there is enough movement in the car's position by opening the boot for it to be re-evaluated and have its attribute re-classified.

[NickM-27]

This would be fairly complicated to change the processing for stationary objects in this way. The object classification feature is designed to represent a way to differentiate objects, not track a live state

[AndyFug]

Understood. I'd read the example given in the docs about a person wearing a helmet/hi-vis and had assumed this was an equivalent case. I guess I mis-interpreted the example and assumed the state was continuously updated.

So you don't think there's any way of achieving what I'm after with the current features? Ie. there is no event or other type of update of an existing object's attributes... even if it moves.

[NickM-27]

right, the current logic looks for the most likely outcome for a given object. So if an object score a certain label at 100% consistently it won't have it's classification changed to another state after. What you're describing is almost like an "object state" classification, something separate we'll need to add support for

[NickM-27]

feel free to make a feature request

[AndyFug]

Thanks. I'll try to find some time to submit the feature request. I can imagine a few scenarios where the feature may be useful, but fully appreciate it may be quite a challenge to develop. :)

Thanks for your comments. That will save me lots of wasted hours trying to make something work that is a non-starter.