Sweet!

I don’t have an external GPU either, just the onboard Intel graphics is what I use now. Also worth mentioning to use integrated graphics your Docker Compose needs:

devices:
      - /dev/dri/renderD128:/dev/dri/renderD128

I’m not using substreams. I have 2 cameras and the motion detection doesn’t stress the CPU too much. If I add more cameras I’d consider using substreams for motion detection to reduce the load.

Your still frames in Home Assistant are the exact problem I was having. If your cameras really do need go2rtc to reduce connections (my wifi camera doesn’t seem to care), you might try changing your Docker container to network_mode: host and see if that fixes it.

Here’s my config. Most of the notations were put there by Frigate and I’ve de-identified everything. Notice at the bottom go2rtc is all commented out, so if I want to add it back in I can just remove the #s. Hope it helps.

mqtt:
  enabled: true
  host: <ip of Home Assistant>
  port: 1883
  topic_prefix: frigate
  client_id: frigate
  user: mqtt username
  password: mqtt password
  stats_interval: 60
  qos: 0

cameras:     # No cameras defined, UI wizard should be used
  baby_cam:
    enabled: true
    friendly_name: Baby Cam
    ffmpeg:
      inputs:
        - path: 
            rtsp://user:pw@<ip-addr>:554/cam/realmonitor?channel=1&subtype=0&unicast=true&proto=Onvif
          roles:
            - detect
            - record
      hwaccel_args: preset-vaapi
    detect:
      enabled: true # <---- disable detection until you have a working camera feed
      width: 1920 # <---- update for your camera's resolution
      height: 1080 # <---- update for your camera's resolution
    record:
      enabled: true
      continuous:
        days: 150
      sync_recordings: true
      alerts:
        retain:
          days: 150
          mode: all
      detections:
        retain:
          days: 150
          mode: all
    snapshots:
      enabled: true
    motion:
      mask: 0.691,0.015,0.693,0.089,0.965,0.093,0.962,0.019
      threshold: 14
      contour_area: 20
      improve_contrast: true
    objects:
      track:
        - person
        - cat
        - dog
        - toothbrush
        - train

  front_cam:
    enabled: true
    friendly_name: Front Cam
    ffmpeg:
      inputs:
        - path: 
            rtsp://user:pw@<ip-addr>:554/cam/realmonitor?channel=1&subtype=0&unicast=true&proto=Onvif
          roles:
            - detect
            - record
      hwaccel_args: preset-vaapi
    detect:
      enabled: true # <---- disable detection until you have a working camera feed
      width: 2688 # <---- update for your camera's resolution
      height: 1512 # <---- update for your camera's resolution
    record:
      enabled: true
      continuous:
        days: 150
      sync_recordings: true
      alerts:
        retain:
          days: 150
          mode: all
      detections:
        retain:
          days: 150
          mode: all
    snapshots:
      enabled: true
    motion:
      mask:
        - 0.765,0.003,0.765,0.047,0.996,0.048,0.992,0.002
        - 0.627,0.998,0.619,0.853,0.649,0.763,0.713,0.69,0.767,0.676,0.819,0.707,0.839,0.766,0.869,0.825,0.889,0.87,0.89,0.956,0.882,1
        - 0.29,0,0.305,0.252,0.786,0.379,1,0.496,0.962,0.237,0.925,0.114,0.879,0
        - 0,0,0,0.33,0.295,0.259,0.289,0
      threshold: 30
      contour_area: 10
      improve_contrast: true
    objects:
      track:
        - person
        - cat
        - dog
        - car
        - bicycle
        - motorcycle
        - airplane
        - boat
        - bird
        - horse
        - sheep
        - cow
        - elephant
        - bear
        - zebra
        - giraffe
        - skis
        - sports ball
        - kite
        - baseball bat
        - skateboard
        - surfboard
        - tennis racket
      filters:
        car:
          mask:
            - 0.308,0.254,0.516,0.363,0.69,0.445,0.769,0.522,0.903,0.614,1,0.507,1,0,0.294,0.003
            - 0,0.381,0.29,0.377,0.284,0,0,0
    zones:
      Main_Zone:
        coordinates: 0,0,0,1,1,1,1,0
        loitering_time: 0

detectors: # <---- add detectors
  ov:
    type: openvino
    device: GPU

model:
  model_type: yolo-generic
  width: 320 # <--- should match the imgsize set during model export
  height: 320 # <--- should match the imgsize set during model export
  input_tensor: nchw
  input_dtype: float
  path: /config/model_cache/yolov9-t-320.onnx
  labelmap_path: /labelmap/coco-80.txt

version: 0.17-0


#go2rtc:
#  streams:
#    front_cam:
#      - ffmpeg:rtsp://user:pw@<ip-addr>:554/cam/realmonitor?channel=1&subtype=0&unicast=true&proto=Onvif
#    baby_cam:
#      - ffmpeg:rtsp://user:pw@<ip-addr>:554/cam/realmonitor?channel=1&subtype=0&unicast=true&proto=Onvif

That CPU has UHD Graphics 750 which is newer than mine which has 730. Should work quite nicely.

Are you using Proxmox, too?

Sounds like LXC is the way to go to pass a Coral through. Not sure why it’s so flaky with the Debian VM.

I’ll keep an eye out for that. So far the Inference Speed is holding stead at 8.47ms.

Are you using OpenVINO with the onboard GPU, or CPU? I think it works with both so you need to make sure it’s using the GPU if possible.

That’s good to hear. That reinforces my suspicion that my problems were caused by passing it through to the virtual machine using Proxmox.

You might be interested in trying to enable the YOLOv9 models. The developer claims they are more accurate, and so far I’m tempted to agree.

You seem a bit more network savvy than me. All I could figure is the Frigate integration (also HACS for me) talks to Frigate and asks it where to get the video from. If go2rtc is enabled in Frigate, the integration tries to stream from go2rtc. Without my Docker stack being in host network mode, it wouldn’t work for me.

With no go2rtc, the Frigate integration asks Frigate where to get the stream, and it’s told to get it from the camera from what I can tell.

All just guesses on my end. Hopefully I don’t sound too sure of myself because I’m not really sure.

1. Portainer is practical, but I switched to Dockge and have been much happier with it. It doesn’t have all of the bells and whistles, but the simplicity makes the workflow much better for me. Give both a try!
2. FreshRSS for RSS feeds, Lubelogger for tracking car (other other things) maintenance, Nginx Proxy Manager for a reverse proxy (or Caddy which is also popular). Whatever you fancy!
3. Not sure.

Regarding domain name, use what you have. It’s super easy to change domain names, and some people do it regularly to take advantage of 1st year sales. Basically all you have to do is transfer your DNS entries to the new domain, and update your reverse proxy entries.

Definitely put everything behind a reverse proxy. I followed this advice so I don’t even have to expose ports using Docker. Everything runs through the reverse proxy, and Dockge makes it trivial add each container to the same network.

Citronella candles and torches don’t repel mosquitos, but they mask the scent of humans. This helps, but not 100%.

DEET doesn’t mask your scent, so mosquitos are still attracted to you. Once they land on your skin, though, they fly away immediately. Very effective.

Propane based traps work very well for collecting large numbers of mosquitos. There isn’t a definitive answer on whether they reduce the local mosquito population over time. The idea is that if large numbers are killed, it will reduce the local breeding.

Thermacell brand mosquito repellers are, IMO, magic. At first I thought they were a gimmick, but my impression of them is very positive. They output a light chemical mist which keeps mosquitoes away. They only work with little to no wind, and they take 15 minutes to warm up and start working.

You want to pick your own MAC? At least you can set it to not be random for a specific network.