[Support]: Object still being detected after disappearing

Describe the problem you are having

My car usually sits in the driveway and is correctly detected as a stationary object by Frigate. However, I noticed this morning that after backing out of the driveway and leaving, the car is still being detected as present.

Here is the car after backing out of the driveway and just before driving away:

Here is what Frigate debug showed right after and continues to show hours later:

I have the street masked off, if that’s relevant:

The car is correctly not being counted as an object in my driveway zone, but is still being counted by the Front camera as being present.

Version

0.10.0-BFECEE9

Frigate config file

mqtt:
  host: 192.168.1.135

detectors:
  coral:
    type: edgetpu
    device: usb

objects:
  track:
    - person
    - car

detect:
  stationary:
    max_frames:
      objects:
        person: 300

record:
  enabled: True
  retain:
    days: 0
  events:
    max_seconds: 300
    retain:
      default: 10
      mode: motion

snapshots:
  enabled: True

ffmpeg:
  hwaccel_args:
    - -c:v
    - h264_v4l2m2m

cameras:
  Front:
    ffmpeg:
      inputs:
        - path: 'rtsp://{FRIGATE_RTSP_PASSWORD}/cam/realmonitor?channel=1&subtype=1'
          roles:
            - detect
    detect:
      width: 704
      height: 480
    motion:
      contour_area: 125
      mask:
        # timestamp
        - 687,20,686,42,514,40,515,19
        # neighbors and road
        - 35,132,76,141,219,93,514,103,703,150,704,68,704,0,527,0,195,0,82,0
    zones:
      driveway:
        coordinates: 570,480,96,480,45,302,494,114,644,146
    record:
      events:
        required_zones:
          - driveway
    snapshots:
      required_zones:
        - driveway

Relevant log output

Nothing relevant in the standard log output.  If I should turn on debug logging, let me know.

FFprobe output from your camera

N/A

Frigate stats

No response

Operating system

Debian

Install method

Docker Compose

Coral version

USB

Network connection

Wired

Camera make and model

Amcrest

Any other information that may be helpful

No response

[blakeblackshear]

Adding the clip as a camera is just an approximation of what happened originally. It’s not going to be exactly the same.

The same logic applies to car objects. There was probably a car briefly detected in the zone that was linked to a car outside the zone in a single frame.

What about my idea to have a configurable parameter for how long an object has to be inside a zone to truly be considered in the zone? If these issues are occurring due to objects very briefly being incorrectly detected (or their bounding boxes very briefly being too large), I would think that setting a minimum threshold of 5-10 frames within a zone would be a much easier way of filtering them out than trying to tweak min_area filters.

Frigate already does something similar when determining if the object is a true positive initially. It takes at least 3 frames to determine that it’s a true positive and the median score of the last 10 detections must be higher than the threshold. This is only happening because there are confirmed true positives (that may not actually be true positives) outside your zone being tracked already. Without that, it would already take multiple frames.

I still think the best way to address this is by improving the object tracking algorithm. A car can’t suddenly jump from one place to another in a single frame.

It may still make sense to require at least a few consecutive frames to consider a zone entered. I will add that to my list.

[blakeblackshear]

It’s possible. The debug view only renders at 5fps, so if your detect fps is higher than that, then some frames aren’t shown.

In the video you showed, there were a lot of very small people getting lost and redetected. If a small shadow or leaf in your zone is recognized as a person (even with a low score) in one frame where someone went missing, it will be linked to the missing object and the it suddenly enters your zone.

The object tracking isn’t that smart yet. All it knows is that it saw 2 people in the frame last time, so if it finds 2 people in the next frame they must be the same 2 people. It doesn’t take into consideration what is a reasonable distance for the object to move, similar colors, or the direction the object was moving. This is why you want to avoid those false positives in the distance even if they are in areas you don’t care about.