So I've been self-hosting my CCTV for about 3 years now and it's always been... not great
First I gave Blue Iris a try which meant I needed a full Windows VM to run it
And it worked - it did the job and recorded stuff and it was fairly OK at motion detection, but damn did it eat the CPU and draw a lot of electricity for no real reason
A few months later I gave Shinobi CCTV a try in Docker and that's what I've been running since
Again, it's mostly fine but the UI is a little clunky and my use case of "24/7 recording that I can easily watch back" was mostly being met, although I had 1 problem
By default Shinobi segments video into 15 minute chunks
So if someone smashes into my car at 14:45:01 then I can't watch that footage until 15:00
Obviously this is a big flaw, so to get around this I changed the segment size to 1 minute
But I have 4 cameras, so this means that over a day I'll now have 5760 clips per day
Sifting through those to find some footage is not fun
Enter Frigate - I'd tried it before but never really gave it a full chance
It's a bit to wrap your head around at first, but once it's up and running it's just a docker-compose.yml for the container and a simple frigate.yml config file
The docs are EXTENSIVE and answered almost every question I had
But there's 1 extra awesome feature I wasn't originally aware of: OpenVINO
OpenVINO is a deep learning model from Intel that apparently runs on my old Broadwell gen Xeon E5-2650v4 CPUs without issue
I've turned it on and enabled object detection and I gotta say, WOW, it's very good
I can go outside with the dog, walk around for a moment and come back in and it'll pick both of us up no problem
So this saved me about £100 seeing as I don't need a Coral compute module (OK I could still get one, but I'm happy for now)
And just to top all of this off, Frigate and Reolink cameras generally don't play too nicely together, yet with support from the docs, mine are working great
Looking at Zabbix, my CPU utilisation for my CCTV server was averaging 10% whilst using Shinobi
Now it's up to 50% but my UPS runtime hasn't really changed so I'm calling that a win
My config is below if it helps anyone trying to get this set up with Reolink cameras
# Disable MQTT because I'm not connecting this to Home Assistant
mqtt:
enabled: false
# Enable 24/7 recording (mode: all means all clips, not just clips with objects in)
# Keep 30 days worth of footage (Frigate automatically deletes the oldest footage once space gets extremely low)
record:
enabled: true
retain:
days: 30
mode: all
# Set the birdseye view to always show a live stream of the cameras
birdseye:
mode: continuous
# Detection area for cameras (all 4 of my Reolink RLC-410 cameras are 2560x1920)
detect:
width: 2560
height: 1920
# Objects to track from /labelmap.txt
objects:
track:
- person
- bicycle
- car
- motorcycle
- bird
- cat
- dog
# Copy and paste from docs to use default OpenVINO
# https://docs.frigate.video/configuration/detectors/#openvino-detector
detectors:
ov:
type: openvino
device: AUTO
model:
path: /openvino-model/ssdlite_mobilenet_v2.xml
model:
width: 300
height: 300
input_tensor: nhwc
input_pixel_format: bgr
labelmap_path: /openvino-model/coco_91cl_bkgr.txt
# Config for each camera
cameras:
c1-side:
ffmpeg:
inputs:
# Record HD stream
- path: http://10.10.8.11/flv?port=1935&app=bcs&stream=channel0_main.bcs&user=admin&password={FRIGATE_C1_PASS}
input_args: preset-http-reolink
roles:
- record
# Use low quality and low FPS stream for object detection
- path: http://10.10.8.11/flv?port=1935&app=bcs&stream=channel0_sub.bcs&user=admin&password={FRIGATE_C1_PASS}
input_args: preset-http-reolink
roles:
- detect
# Record audio
output_args:
record: preset-record-generic-audio-copy
c2-garden:
ffmpeg:
inputs:
- path: http://10.10.8.12/flv?port=1935&app=bcs&stream=channel0_main.bcs&user=admin&password={FRIGATE_C2_PASS}
input_args: preset-http-reolink
roles:
- record
- path: http://10.10.8.12/flv?port=1935&app=bcs&stream=channel0_sub.bcs&user=admin&password={FRIGATE_C2_PASS}
input_args: preset-http-reolink
roles:
- detect
output_args:
record: preset-record-generic-audio-copy
c3-garage:
ffmpeg:
inputs:
- path: http://10.10.8.13/flv?port=1935&app=bcs&stream=channel0_main.bcs&user=admin&password={FRIGATE_C3_PASS}
input_args: preset-http-reolink
roles:
- record
- path: http://10.10.8.13/flv?port=1935&app=bcs&stream=channel0_sub.bcs&user=admin&password={FRIGATE_C3_PASS}
input_args: preset-http-reolink
roles:
- detect
output_args:
record: preset-record-generic-audio-copy
c4-front:
ffmpeg:
inputs:
- path: http://10.10.8.14/flv?port=1935&app=bcs&stream=channel0_main.bcs&user=admin&password={FRIGATE_C4_PASS}
input_args: preset-http-reolink
roles:
- record
- path: http://10.10.8.14/flv?port=1935&app=bcs&stream=channel0_sub.bcs&user=admin&password={FRIGATE_C4_PASS}
input_args: preset-http-reolink
roles:
- detect
output_args:
record: preset-record-generic-audio-copy
Regardless of what cameras you choose, please ensure you VLAN and firewall them off - these cameras effectively run a Linux distro and should not be trusted or accessible
For example, my Reolink cameras can access NTP and DNS just so their clocks are correct
They can't access anything else on the network
The CCTV VM sits on the same network as the cameras and has host firewall rules to deny access from the cameras
Frigate just connects to each camera's stream and does its magic from there
version: "3.9"
services:
frigate:
container_name: frigate
privileged: true # this may not be necessary for all setups
restart: unless-stopped
image: ghcr.io/blakeblackshear/frigate:stable
shm_size: "512mb" # update for your cameras based on calculation above
volumes:
- /etc/localtime:/etc/localtime:ro
- /opt/dockervolumes/frigate/config/config.yml:/config/config.yml
- /mnt/cctv/frigate:/media/frigate
- type: tmpfs # Optional: 1GB of memory, reduces SSD/SD Card wear
target: /tmp/cache
tmpfs:
size: 1000000000
ports:
- "8554:8554" # RTSP feeds
- "8555:8555/tcp" # WebRTC over tcp
- "8555:8555/udp" # WebRTC over udp
environment:
FRIGATE_C1_PASS: ${FRIGATE_C1_PASS}
FRIGATE_C2_PASS: ${FRIGATE_C2_PASS}
FRIGATE_C3_PASS: ${FRIGATE_C3_PASS}
FRIGATE_C4_PASS: ${FRIGATE_C4_PASS}
labels:
- traefik.enable=true
- traefik.http.routers.cctv.rule=Host(`cctv.${DOMAIN}`)
- traefik.http.routers.cctv.entrypoints=websecure
- traefik.http.routers.cctv.tls.certresolver=cloudflare
- traefik.http.services.cctv.loadbalancer.server.port=5000
networks:
- proxy
networks:
proxy:
external: true
Blue Iris user here, I heard good things about Frigate before, but like you, I never tried it because it seems hard to set up and configure. But since I also have Reolink cameras and with your docker compose.yml file, I might give it a shot.
First I gave Blue Iris a try which meant I needed a full Windows VM to run it
"And it worked—it did the job, recorded stuff, and was fairly okay at motion detection. However, it consumed a significant amount of CPU and drew a lot of electricity for no real reason. You didn't optimize it.
For me, with 4 4k Cameras + more 1080p cameras, it uses 1.5% to 3% when nobody is viewing the cameras. Of course, it will go up if somebody is viewing the cameras, as it will need to transcode to view it on the web. I wanted to try other alternatives as well on Linux, but I still preferred Blue Iris at the end.
You can find information on optimizing Blue Iris's CPU usage in this link: Optimizing Blue Iris's CPU Usage.
One thing to mention is that you're currently running detection on 2560x1920 but your reolink ext stream is going to be something like 896x672 so you're using cpu to scale that stream up to the higher resolution. You may find lower cpu usage by running detect at that streams specific resolution
One thing that Shinobi could do was integrate well with the detectors on cam, so if your cam had person/vehicle detection or line crossing etc you could use that. That worked well.
But I too went from Shinobi to Frigate and I love it. The integration with HA is very good.
Have been running my own NVR for decades, most of it with the motion project. Recently started switching things over to Frigate and I am impressed with it as well. The Corals are worth the money. I run two of them on my setup (6 to 8 1080 cams at any given time) and quite pleased with it.
The default models are great at detecting humans, but not so much cars/wildlife. I definitely plan to try and expand on that in the near future.