[Config Support]: Lower min_score for stationary objects?

[rsteckler]

Describe the problem you are having

I often have a false positive pop up that gets detected for 10 seconds and becomes stationary.
The problem is that this remains detected for a very long time because I think the detection every 50 frames for a stationary object is returning very low scores, under min_score, which are then being discarded. That means they aren't contributing to the overall score for the object and the object remains above the threshold for hours. It seems to unstick once the bounding box moves enough to be considered active again, or if there's incidental movement in the area, like an actual person walking past.

Here's an example:

That firepit was detected as a person for an hour and I suspect there was an hour-long string of scores under the min_score that probably should have contributed to understanding it's not actually a person.

Hope this makes sense? it might be good to add a separate min_score specifically for continued detection of stationary objects, so we can lower that to zero so that false positives can be cleared more quickly. I may also just be misunderstanding how this works.

Version

0.13.0-C35C7DA

Frigate config file

mqtt:
  host: homestack-mosquitto-service

detectors:
  coral:
    type: edgetpu
    device: usb

database:
  path: /config/database.db

model:
  path: plus:https://73b30a3d84dec0b38affc6f39ca63475

birdseye:
  # Optional: Enable birdseye view (default: shown below)
  enabled: true
  # Optional: Width of the output resolution (default: shown below)
  width: 1920
  # Optional: Height of the output resolution (default: shown below)
  height: 1080
  # Optional: Encoding quality of the mpeg1 feed (default: shown below)
  # 1 is the highest quality, and 31 is the lowest. Lower quality feeds utilize less CPU resources.
  quality: 8
  # Optional: Mode of the view. Available options are: objects, motion, and continuous
  #   objects - cameras are included if they have had a tracked object within the last 30 seconds
  #   motion - cameras are included if motion was detected in the last 30 seconds
  #   continuous - all cameras are included always
  mode: motion

record:
  enabled: True
  events:
    retain:
      mode: active_objects
      default: 10 


motion:
  # Optional: The threshold passed to cv2.threshold to determine if a pixel is different enough to be counted as motion. (default: shown below)
  # Increasing this value will make motion detection less sensitive and decreasing it will make motion detection more sensitive.
  # The value should be between 1 and 255.
  threshold: 35

ffmpeg:
  hwaccel_args: preset-vaapi

go2rtc:
  streams:
    camera_pool:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_rear_patio:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_driveway:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_front_gate_driveway:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_front_gate_walkway:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_garage:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_balcony:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_front_yard:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_casita:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_front_door:
      - rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
    camera_familyroom:
      - rtsp:https://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=0
    camera_familyroom_pool:
      - rtsp:https://admin:[email protected]:554/cam/realmonitor?channel=1&subtype=0

cameras:
  #    debug_test:
  #  ffmpeg:
  #    inputs:
  #      - path: /media/frigate/dog2.mp4
  #        input_args: -re -stream_loop -1 -fflags +genpts
  #        roles:
  #          - detect
  #  detect:
  #    height: 2688
  #    width: 1520
  #    fps: 5
  #  rtmp:
  #    enabled: False
  #  snapshots:
  #    enabled: True
  #    timestamp: False
  #    bounding_box: True
  #    crop: False
  #    retain:
  #      default: 365
  camera_pool:
    mqtt:
      enabled: true
      timestamp: false
      bounding_box: false
      crop: true
      quality: 100
      height: 1080
    ffmpeg:
      inputs:
      - path: rtsp:https://admin:[email protected]/cam/realmonitor?channel=1&subtype=2
        roles:
        - detect
    detect:
      width: 1920
      height: 1080
    snapshots:
      enabled: true
      timestamp: false
      bounding_box: true
      crop: false
      retain:
        default: 365
        objects:
          person: 50
objects:
  track:
  - person
  - dog
  - cat
  - car
  - face
  - package
  - amazon
  - fedex
  - ups
  - license_plate
  filters:
    dog:
      min_score: .7
      threshold: .9
    cat:
      min_score: .65
      threshold: .8
    face:
      min_score: .7
      threshold: .7
    package:
      min_score: .65
      threshold: .75
    license_plate:
      min_score: .6
      threshold: .7
    amazon:
      min_score: .75
      threshold: .7
    ups:
      min_score: .75
      threshold: .7
    fedex:
      min_score: .75
      threshold: .7
    person:
      min_score: .65
      threshold: .85
    car:
      min_score: .65
      threshold: .85

Relevant log output

NA

Frigate stats

No response

Operating system

Other Linux

Install method

Docker Compose

Coral version

USB

Any other information that may be helpful

No response

[NickM-27]

which are then being discarded. That means they aren't contributing to the overall score for the object and the object remains above the threshold for hours.

that's not how that works. If the score is below min_score then that means the detection is discarded which means the object is considered missing in the frame (ie not detected). For the object to continue to be stationary and ongoing it must be detected which means the score must be above min_score. Lowering min_score for stationary objects would just make the stationary false positives stick around longer

[rsteckler]

Thanks for explaining, but I think maybe the docs are wrong then?

I'm assuming that:

1. after an object is considered stationary, that the detection is run every intervalth frame (default 50 or 10 seconds).
2. It's considered "gone" if the computed detection falls below the threshold. The computed detection is the median of all previous detections.
3. If any given detection is below min_score, the data point is thrown away and not considered in computed detection score that gets compared to threshold.

Is my third assumption incorrect? I'm getting that belief from this page that shows that the 0.55 frame isn't included in any of the following frames to compute the overall score.

[NickM-27]

The docs you are referencing are having to do with when an object is considered a true positive and an event is saved, not what happens to it after it is detected and being tracked as a true positive.

For an object to be gone it must not be detected for multiple frames. If an object's score is detected as below min_score then that detection is thrown out and it means the object is considered not detected for that frame