[Config Support]: Problem detecting people further "away"

Describe the problem you are having

hello everyone, I use a frigate with 4 cameras. Everything works fine in 3 of them, while in the fourth one (the one with the best definition, called “Giardino”) I have a little problem with detecting people. In practice, the view is very wide, and people are only detected in the red part, while they are rarely seen in the yellow part, being further away and therefore smaller. I’ve read all the documentation, but I’m a newbie and I don’t quite understand how I can fix this. Thank you

Version

0.11.1-2EADA21

Frigate config file

mqtt:
  host: 192.168.1.56
  user: xxx
  password: xxx
  topic_prefix: frigate
  client_id: frigate

cameras:
  Ingresso:
    ffmpeg:
      inputs:
        - path: rtsp://xxx:xxx@192.168.1.65:554/live2.sdp
          roles:
            - detect
        - path: rtsp://xxx:xxx@192.168.1.65/live1.sdp
          roles:
            - record
      hwaccel_args: -c:v h264_qsv
    rtmp:
      enabled: false
    detect:
      width: 480
      height: 272
      fps: 7
    motion:
      mask:
        - 265,87,311,92,319,55,269,51
    objects:
      track:
        - person
        - bicycle
        - dog
        - cat
        - car
    snapshots:
      enabled: true
      bounding_box: true
  Garage:
    ffmpeg:
      inputs:
        - path: rtsp://xxx:xxx@192.168.1.209:554/cam/realmonitor?channel=1&subtype=1&unicast=true&proto=Onvif
          roles:
            - detect
        - path: rtsp://xxx:xxx@192.168.1.209/cam/realmonitor?channel=1&subtype=00&authbasic=YWRtaW46TDI4NUYxNjQ=
          roles:
            - record
      hwaccel_args: -c:v h264_qsv
    rtmp:
      enabled: false
    detect:
      width: 640
      height: 480
      fps: 15
    objects:
      track:
        - person
        - bicycle
        - dog
        - cat
        - car
    snapshots:
      enabled: true
      bounding_box: true
  Giardino:
    ffmpeg:
      inputs:
        - path: rtsp://xxx:xxx@192.168.1.77:554/h264Preview_01_sub
          roles:
            - detect
        - path: rtsp://xxx:xxx@192.168.1.77:554/h264Preview_01_main
          roles:
            - record
    rtmp:
      enabled: false
    detect:
       width: 1536
       height: 576
       fps: 30
    motion:
      mask:
        - 0,0,388,0,0,504
        - 1536,0,1536,322,1384,440,1067,0
    objects:
      track:
        - person
        - dog
        - cat
        - car
    snapshots:
      enabled: true
      bounding_box: true
  Sala:
    ffmpeg:
      inputs:
        - path: rtsp://xxx:xxx@192.168.1.64/Streaming/Channels/2
          roles:
            - detect
        - path: rtsp://xxx:xxx@192.168.1.64/HighResolutionVideo
          roles:
            - record
      hwaccel_args: -c:v h264_qsv
    rtmp:
      enabled: false
    detect:
      width: 352
      height: 240
      fps: 15
    objects:
      track:
        - person
        - dog
        - cat
    snapshots:
      enabled: true
      bounding_box: true
birdseye:
  enabled: True
  mode: continuous
detectors:
  coral:
    type: edgetpu
    device: usb
record:
  enabled: True
  events:
    retain:
      default: 10
database:
  path: /media/frigate/db/frigate.db

Relevant log output

nothing

Frigate stats

No response

Operating system

Proxmox

Install method

Docker Compose

Coral version

USB

Any other information that may be helpful

No response

[blakeblackshear]

A higher detect resolution could help here too.

[NickM-27]

The issue is that sub-stream is used for both motion detection and object recognition, while it would be wiser to detect motion on sub, and then crop the main stream without analyzing it and forward it to coral for object recognition.

It’s not that simple. Not only would you be decoding both streams which is a lot of work, the decoding of different resolutions can have differing delay introduced. The streams from the cameras themselves are also often somewhat out of sync.

This is of course why the docs recommend cameras with configurable sub streams.

More CPU intensive, yes, but I’d better have some extra CPU usage and a working system.

If you don’t care about cpu usage then just use the high res stream and set the resolution in detect -> width / height to downsize it to 1080p.