EdgeTPU found but not working

[thehijacker]

Hello,

Giving Viseron a try and hopefuly it can replay Frigate and Blue Iris for me. I have Google Coral USB version connected, visible on host and docker container and also detected:

21. 12. 2022 9:05:29
[2022-12-21 09:05:29] [DEBUG   ] [viseron.components.edgetpu] - Available devices: [{'type': 'usb', 'path': '/sys/bus/usb/devices/4-2'}]

But then I start getting:

21. 12. 2022 9:07:30
[2022-12-21 09:07:30] [ERROR   ] [viseron.components.edgetpu] - Error when trying to load EdgeTPU: Failed to load delegate from libedgetpu.so.1
21. 12. 2022 9:07:30
21. 12. 2022 9:07:30
[2022-12-21 09:07:30] [ERROR   ] [viseron.components] - Domain object_detector for component edgetpu is not ready. Retrying in 40 seconds. Error: 

Cameras are all streams from Blue Iris in MJPEG format. My config file looks like:

ffmpeg:
  camera:
    vhod:
      name: Vhod
      host: 192.168.28.59
      port: 81
      path: !secret vhod
      stream_format: mjpeg
      fps: 5
      recorder:
        idle_timeout: 1
        codec: h264

    garaza:
      name: Garaza
      host: 192.168.28.59
      port: 81
      path: !secret garaza
      stream_format: mjpeg
      fps: 5
      recorder:
        idle_timeout: 1
        codec: h264

    nadstresek:
      name: Nadstresek
      host: 192.168.28.59
      port: 81
      path: !secret nadstresek
      stream_format: mjpeg
      fps: 5
      recorder:
        idle_timeout: 1
        codec: h264

    spredaj:
      name: Spredaj
      host: 192.168.28.59
      port: 81
      path: !secret spredaj
      stream_format: mjpeg
      fps: 5
      recorder:
        idle_timeout: 1
        codec: h264

    zadaj:
      name: Zadaj
      host: 192.168.28.59
      port: 81
      path: !secret zadaj
      stream_format: mjpeg
      fps: 5
      recorder:
        idle_timeout: 1
        codec: h264

    balkon:
      name: Balkon
      host: 192.168.28.59
      port: 81
      path: !secret balkon
      stream_format: mjpeg
      fps: 5
      recorder:
        idle_timeout: 1
        codec: h264

#darknet:
#  object_detector:
#    cameras:
#      vhod:
#        fps: 1
#        scan_on_motion_only: false
#        labels:
#          - label: person
#            confidence: 0.75
#            trigger_recorder: true

# You can also use motion detection
mog2:
  motion_detector:
    cameras:
      vhod:
        fps: 1
      garaza:
        fps: 1
      spredaj:
        fps: 1
      zadaj:
        fps: 1
      balkon:
        fps: 1
      nadstresek:
        fps: 1

nvr:
  vhod:
  garaza:
  spredaj:
  zadaj:
  balkon:
  nadstresek:

edgetpu:
  object_detector:
    device: usb
    cameras:
      vhod:
        fps: 1
        labels:
          - label: car
            confidence: 0.8
          - label: person
            confidence: 0.8
      garaza:
        fps: 1
        labels:
          - label: car
            confidence: 0.8
          - label: person
            confidence: 0.8
      spredaj:
        fps: 1
        labels:
          - label: car
            confidence: 0.8
          - label: person
            confidence: 0.8
      zadaj:
        fps: 1
        labels:
          - label: car
            confidence: 0.8
          - label: person
            confidence: 0.8
      nadstresek:
        fps: 1
        labels:
          - label: car
            confidence: 0.8
          - label: person
            confidence: 0.8

#  image_classification:
#    device: usb
#    cameras:
#      vhod:
#      garaza:
#      spredaj:
#      zadaj:
#      nadstresek:
#    labels:
#      - dog

webserver:
#  port: 8888

logger:
  default_level: debug

And my docker compose:

  viseron:
    image: roflcoopter/viseron:latest
    privileged: true
    container_name: viseron
    volumes:
      - "/mnt/usb_3/Viseron/recordings:/recordings"
      - "/mnt/usb_3/Viseron/config:/config"
      - "/etc/localtime:/etc/localtime:ro"
      - type: tmpfs
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    devices:
      - /dev/bus/usb:/dev/bus/usb
      - /dev/dri/renderD128
    restart: unless-stopped
    ports:
      - "8200:8888"
    deploy:
      resources:
        limits:
          memory: 1024M
    environment:
#      - PUID=1000
#      - PGID=1000
      - TZ=Europe/Ljubljana

GUI opens but it doesn't display cameras. In log I can see:

21. 12. 2022 9:07:36
[2022-12-21 09:07:36] [WARNING ] [tornado.access] - 404 GET /api/v1/camera/nadstresek/snapshot?rand=fhd91&width=568 (192.168.28.58) 14.90ms
21. 12. 2022 9:07:36
[2022-12-21 09:07:36] [DEBUG   ] [viseron.components.webserver.api] - Routing to CameraAPIHandler.get_snapshot(*args=[], **kwargs={'camera_identifier': b'garaza'})
21. 12. 2022 9:07:36
[2022-12-21 09:07:36] [WARNING ] [tornado.access] - 404 GET /api/v1/camera/garaza/snapshot?rand=5noi2&width=568 (192.168.28.58) 2.49ms
21. 12. 2022 9:07:36
[2022-12-21 09:07:36] [DEBUG   ] [viseron.components.webserver.api] - Routing to CameraAPIHandler.get_snapshot(*args=[], **kwargs={'camera_identifier': b'spredaj'})
21. 12. 2022 9:07:36
[2022-12-21 09:07:36] [WARNING ] [tornado.access] - 404 GET /api/v1/camera/spredaj/snapshot?rand=o0gfrg&width=568 (192.168.28.58) 2.98ms

What else can I try to make Google Coral work? This is fist step. Next step will be actualy recording on detection. And MQTT so I can see this events in Home Assistant. But first I need person and car detected using USB Google Coral working :).

Thank you.

[roflcoopter]

Sorry you are having issues, will investigate asap!

[roflcoopter]

You dont happen to have some other container running at the same time using the EdgeTPU?

[thehijacker]

Good question. Frigate was not running, but it seems Coral is not available anymore. I could see it under lsusb but it did not work on Frigate either anymore. I tried to reset USB from SSH, but it failed and now lsusb doesn't even see it :). When I get home I will replug it back to the server and try again, making sure that Frigate is not running. Will write here what I found.

[roflcoopter]

Great, i just tested on my machine and it works there, so atleast we know the builds are not completely busted