I've now managed to get the stream of my HiWatch Series, model HWI- T641H-Z 2.8-12 mm by accessing via browser to rtsp://USERNAME:PASSWORD@192.168.30.50:554/Streaming/Channels/101. Working nice. But on its own, very unusful unless I get storage and image detection going. So, I need something like Frigate.

I believe the Frigate docs suggest running in a docker container installed on a VM. Very convienient as I have Proxmox (installed on a optiples 7070 micropc with 32MB ram) running a VM on which i have a few containers. The idea is to mount a "frigate" dataset created on my Truenas server on Proxmox which is then mounted to the VM. Lots of layers but hoping it should be ok.

I heard that the Google Coral TPU is not sold anymore. Is this going to be a problem for Frigate image recognition? Are there alternatives to the Coral TPU?

In my experience frigate is quite CPU intensive. I have 10 cameras and my old I7-8700 would die trying to have frigate process them all.

I opted for a different approach with LazyNVR (https://codeberg.org/LazyNVR/lazynvr-sources) leveraging the capability of the camera to perform motion detection and video recording directly. This offloaded my CPU and now I can record all 10 cameras on a OPi3.

You can still get the Google Coral (both the USB dongle and M.2/PCIe versions), but Google has effectively abandoned the official drivers. You'll need to use community-maintained DKMS drivers if you go with the PCIe version on newer Linux kernels.

If you want a modern alternative, look into the Hailo-8 or Hailo-8L M.2 AI accelerators. They plug right into an open M.2 slot and natively integrate into recent versions of Frigate. Keep in mind that these accelerators ONLY process object detection inference - they do not handle video decoding.

For video decoding, your OptiPlex 7070's Intel CPU has Quick Sync Video (QSV). You can pass /dev/dri through to your container, and Intel iGPU hardware acceleration will handle decoding the RTSP streams with minimal CPU overhead.

If you prefer an all-in-one approach, an Intel or Nvidia GPU can handle both video decoding and object detection (via OpenVINO or TensorRT), but a Micro PC limits your physical expansion options.

Storing recordings on a TrueNAS share over NFS or SMB works just fine. Mount the NFS share directly in Proxmox or the VM host, then pass the directory through to your Docker container.

I have the same setup. Run ov on the igpu, works great.

Regarding Google Coral, from the docs: https://docs.frigate.video/frigate/hardware

The Coral is no longer recommended for new Frigate installations, except in deployments with particularly low power requirements or hardware incapable of utilizing alternative AI accelerators for object detection. Instead, we suggest using one of the numerous other supported object detectors. Frigate will continue to provide support for the Coral TPU for as long as practicably possible given its still one of the most power-efficient devices for executing object detection models.

Personally, I'd recommend OpenVINO. It works on iGPUs, and if that's not possible it's quite performant on CPUs. I currently run a Google Coral and the Linux driver is no longer officially maintained meaning it breaks every now and then after a kernel update, next time it breaks I'm going with OpenVINO.

Also, I'd install Docker in an LXC for Frigate, it makes passing the iGPU through to it a lot easier.

Have an i3 8xxx running 7 cams. Avg 15% detection CPU.
Took a while to get my config right.

My complaint isn’t that frigate is inefficient it’s that it’s way too easy to make it inefficient.

The Frigate docker image is generally nasty. It's humongous (5.5 GB), bundles and runs a ton of different things whether you use them or not, uses s6 as init and supervisor which is a piece of crap, and it cannot be secured – it won't run as a non-privileged user, it won't drop caps, you can't make it read-only because some genius configured nginx to put temporary files in with the app files, it conveniently includes apt so the attacker can install anything they might want inside the container, and in fact recommends running in privileged mode(!).

I think it's the most security-hostile docker image I have ever seen.

I've tried LazyNVR but I couldn't figure out how to make it work with my cameras. There seem to be no real instructions. I don't even get how it's supposed to find the cameras, there's no place to put an IP.

Feels like nextcloud....

Good! I was probably doing something wrong, but in any case seems more logical to leverage the cameras internal detection.

Yeah Frigate devs claim it can't be done, but it works in other projects like Shinobi.

But if has all these issues, why is it so popular, and more importantly, hoe come the developers are not improving the image? Is it because its too complex?

It's popular because it works. You don't get a lot of choice in the NVR area, most of the other projects are kind of ass too. I've been trying lots of them and so far Frigate and Shinobi are the only ones that you can get to show your camera streams, and detecting & recording actually works.

I also don't think many people care about the security that much. Either that or they run the container like I do in a completely isolated VLAN with the cameras. Honestly, if it wasn't for the obscene amounts of RAM, CPU and storage it wastes for no good reason I wouldn't care that much about the security either.

hoe come the developers are not improving the image?

Lots of developers are clueless about Docker. They are used to running their stuff on the metal and piling everything together and they don't even consider they should clean things up before they ship a docker image.

Gramps Web for example is another humongous image, and it's humongous because apparently the developer can't be arsed to separate the building stage from the production stage in their Dockerfile, so they're shipping all the build toolchains. People have pointed this out to them and they're like "eh whatever".

Acronyms, initialisms, abbreviations, contractions, and other phrases which expand to something larger, that I've seen in this thread:

6 acronyms in this thread; the most compressed thread commented on today has 10 acronyms.

[Thread #104 for this comm, first seen 15th Sep 2026, 06:30]
[FAQ] [Full list] [Contact] [Source code]

Nah, not even close. Nextcloud's attempt to throw in the kitchen sink is what makes it bloated and inefficiant. Frigate on the other hand isn't like that. The bloat there comes from misconfigured pipelines for your camera feeds. You need some time to dial them in but once you do, it's literally years between maintenance and very little overhead.
I have 10 cameras and my CPU barely breaks 50% utilization. I live on a busy street with lots of traffic.I can literally see a statistical trend when people go to work, go out, etc, based on cpu usage alone. That being said, not all feeds run all types of analysis. Some have only motion, some have everything, even audio.