TL;DR

1
2
3
4
5

mkdir -p /opt/frigate/config /opt/frigate/media
cd /opt/frigate
# write docker-compose.yml and config/config.yml (below)
docker compose up -d
docker compose logs -f frigate

Authenticated UI at https://<host>:8971. On first run it prints the generated admin password in the log.

Frigate is an NVR with real-time object detection (person, car, animal…) for IP cameras. It runs happily in a container: give it a docker-compose.yml, a config.yml and your cameras’ RTSP URLs. The rest is tuning shared memory and hardware acceleration.

docker-compose.yml

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27

services:
  frigate:
    container_name: frigate
    image: ghcr.io/blakeblackshear/frigate:stable
    restart: unless-stopped
    stop_grace_period: 30s
    shm_size: "256mb"            # raise it per camera count and resolution
    devices:
      - /dev/dri/renderD128      # iGPU (VAAPI) for decoding; optional
      # - /dev/bus/usb:/dev/bus/usb   # Coral USB
      # - /dev/apex_0:/dev/apex_0     # Coral PCIe
    volumes:
      - /etc/localtime:/etc/localtime:ro
      - ./config:/config
      - ./media:/media/frigate
      - type: tmpfs
        target: /tmp/cache
        tmpfs:
          size: 1000000000
    ports:
      - "8971:8971"     # authenticated UI (use this one)
      - "8554:8554"     # RTSP restream (go2rtc)
      - "8555:8555/tcp" # WebRTC
      - "8555:8555/udp"
    environment:
      FRIGATE_RTSP_USER: "${FRIGATE_RTSP_USER}"
      FRIGATE_RTSP_PASSWORD: "${FRIGATE_RTSP_PASSWORD}"

Camera credentials live in a .env next to the compose file (never in git):

1
2
3

# /opt/frigate/.env
FRIGATE_RTSP_USER=viewer
FRIGATE_RTSP_PASSWORD=change-me

config/config.yml

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

mqtt:
  enabled: false        # or point it at a mosquitto broker for Home Assistant

detectors:
  cpu1:
    type: cpu           # switch to edgetpu (Coral) or tensorrt (NVIDIA)

ffmpeg:
  hwaccel_args: preset-vaapi   # Intel/AMD iGPU; NVIDIA: preset-nvidia

go2rtc:
  streams:
    door1:
      - "rtsp://{FRIGATE_RTSP_USER}:{FRIGATE_RTSP_PASSWORD}@camera-ip:554/stream1"

cameras:
  door1:
    ffmpeg:
      inputs:
        - path: rtsp://127.0.0.1:8554/door1   # single connection via go2rtc
          roles: [detect, record]
    detect:
      width: 1280
      height: 720
      fps: 5
    objects:
      track: [person, car, dog]
    record:
      enabled: true
      retain:
        days: 7
    snapshots:
      enabled: true

Add more cameras by repeating the go2rtc.streams + cameras blocks. Reload with docker compose restart frigate.

What each piece does

PiecePurpose
shm_sizeFrame buffers in /dev/shm. Too small = cameras crash.
go2rtcRepackages the camera RTSP and serves it on :8554. Frigate reads from there to detect and record over a single connection to the camera.
hwaccel_argsOffloads video decode to the GPU/iGPU. Without it the CPU melts.
detectorsWhere inference runs: cpu (slow), edgetpu (Coral) or tensorrt (NVIDIA).
Port 8971UI with login. This is the one you expose.
record / snapshotsWrite to ./media. Retention decides how much disk you burn.

Traps

  • Small shm_size → cameras that restart themselves. Rough per-camera rule: (width*height*1.5*9 + 270480) / 1048576 MB; sum all cameras and leave a base. A few 1080p cameras will blow past 64–128 MB.
  • Do not expose port 5000. It’s the unauthenticated internal API. The public UI is 8971 (with login). Keep 5000 on loopback / an internal network only.
  • Use the go2rtc restream. Putting the camera RTSP straight into both detect and record opens two connections to the sensor; many cheap cameras can’t take it. Detect and record from rtsp://127.0.0.1:8554/<cam>.
  • Detect at low resolution. detect at 1280×720 and 5 fps is plenty; record the high stream separately if your camera has two.
  • Hardware acceleration or 100% CPU. preset-vaapi (Intel/AMD) or preset-nvidia; the /dev/dri device (or the NVIDIA runtime) must be mapped in.
  • A detector isn’t really optional. The cpu detector is fine to smoke-test, but past one or two cameras you want Coral or TensorRT.

See also