mydocker/frigate/README.md

Frigate - AI-Powered NVR

Real-time object detection NVR system. 8 cameras, detects persons, cars, dogs and cats. Integrates with Home Assistant via MQTT.

Access

Endpoint	Details
Web UI	http://acemagic:5000 (local only, no Traefik)
API	http://acemagic:5000/api/
go2rtc restream	http://acemagic:8554/ (RTSP)

Configuration

• Config version: 0.17-0 (config.yml)
• RAM limit: 6 GB
• MQTT: 192.168.1.132:1883 (Mosquitto), topic prefix frigate

Hardware Acceleration

Component	Hardware	Purpose
Google Coral TPU	USB (device: usb)	AI object detection
AMD GPU radeonsi	VAAPI	Video decoding (hwaccel)

Cameras

Name	Model	IP	Tracked Objects	Zones
reolink_1	Reolink	192.168.1.101	person, car, dog	parking, driveway, r1_my-garden
reolink_2	Reolink	192.168.1.62	person, dog	my_sidewalk
reolink_3	Reolink	192.168.1.150	person, car, dog	terrace, kennel, my-garden
reolink_4	Reolink	192.168.1.168	person, car, dog	main_area, entrance
reolink_5	Reolink	192.168.1.172	person, dog	terrace
hikvision_2	Hikvision	192.168.1.65	person, car, dog	my-garden
hikvision_4	Hikvision (intercom)	192.168.1.127	person	street-left, robo, at-gate
mercusys_mc200	Mercusys (indoor)	192.168.20.33	person, cat	-

All streams are rerouted through built-in go2rtc (RTSP restream) to save RAM - Frigate pulls one stream from go2rtc instead of directly from each camera.

Retention

Type	Duration
Snapshots (persons)	3650 days (10 years)
Snapshots (cars)	180 days
Snapshots (default)	365 days
Recordings - alerts	30 days
Recordings - detections	90 days
Continuous recordings (hikvision_4)	90 days
Recordings with motion	7 days (reolink_1 only)

Snapshot quality: 90%, with clean copies (clean_copy: true).

Architecture

Camera IP (RTSP)
  └── go2rtc (internal restream)
        └── Frigate detection (Coral TPU, 640x360 @ 7-10fps)
              ├── MQTT → Mosquitto → Home Assistant (events, binary sensors)
              ├── recordings → /media/evo2t/frigate/
              └── snapshots → frigate.db + /media/evo2t/frigate/clips/

Key Configuration Details

• Detection: 640x360, 7-10 fps (sub-stream), stationary object suppression
• Semantic search: enabled (model small)
• Birdseye: 1920x1080, continuous mode
• Motion masks per camera (eliminates false alarms from shadows, road signs)

Useful Commands

# Follow logs
docker compose logs -f frigate

# Restart
docker compose restart frigate

# Check RAM usage
docker stats frigate --no-stream

# Watch Frigate MQTT events
mosquitto_sub -h 192.168.1.132 -u mosquitto -P <password> -t 'frigate/#' -v

# Clean old recordings manually (auto-handled by Frigate)
# Use UI: http://acemagic:5000 → Storage