Hi Jim. Thank you for referring me to this video in your response to my comment on your UniFi Controller on Docker video! (Shout-out to Lawrence Systems for referring to your video in his recent UniFi video) I find your videos very useful and informative. I thoroughly enjoy how you go in-depth with explaining the compose .yaml file. I would also like to thank you for bringing up the Coral TPU. I never realized such a thing existed. You have earned my Like and Subscription, and I hope you earn more from others, as well
I really appreciate your feedback, thank you. Important to note that since the video an integrated GPU is now supported, thus the TPU isn't strictly required.
@@Jims-Garage Good to know. I will likely be saving my iGPU resources for a display in case I need to do some troubleshooting. With the TPU being so small and relatively inexpensive, as well as being another thing to experiment with, I will likely go with the TPU route
Just a note: if you don't want e.g.: the iGPU to be exclusive to frigate (to use for decoding/transcoding) use a docker LXC instead of a full VM. I have a miniITX system that has the iGPU used by both proxmox (for graphical console access) and frigate on a docker that is in a LXC. Also to add: not all m.2 e-key pcie wifi slots support two devices, so some might only see a single device.
If you use a computer with an intel cpu, you can also use the openvino model, and vaapi for acceleration. That model seems to be better for detecting people, and does not need a coral, and isn't as cpu heavy as using just cpu.
the detect feed should be low res and low fps. 5 fps and ~640x480 are enough for detection, and is way less load on the TPU and CPU the detect and the record stream URLs can be different
Is the MQTT server only necessary for using Frigate with Home Assistant? Is it needed otherwise? I was going to run it directly, without HA integration. Great video, btw!
Hi Jim, all your project are really cool, i got some problems with frigate, in the future I'll buy coral hw, but for the moment i just want to see the web ui with a camera, to take a look at the interface, when i start the container i can open the web ui, but it still load, and five me only the top bar with nothing else, anyone know why? I don't use exactly your same config.
Hello, I habe got a question. Would it be possible to run all the software including proxmox & Frigate on a RP4 while having the about the same performance? This is intended to get a mobile AI system. Would there be any bottlenecks? We would run the system on a SSD instead of using SD cards of course. Thanks for the nice video. I appreciate any comments on this.
Hey, good question, I haven't ever done it, but... Proxmox has a small performance impact of around a few percent, so not much. Docker is also very lightweight, but you could run binaries. The most important thing would be to use a coral USB, there's no way the Pi's CPU could do AI detection. With a coral AI it would likely be very capable for CCTV.
Not sure if you have settled on a RPi4 but I would seriously consider an RPi5 instead since it has native m.2 compatability now with a pi hat. It would make the system very viable without any workarounds messing with performance in anyway. Would still need a coral. But it will be a much more viable system for longer and will survive many updates.
Hello very great video. Hello ! I have my home assistant on a rasberry 4. I have now bought myself a Google Coral Edge TPU USB Accelerator. How do I install it in the raspberry? Is it just to insert it into usb and run or do I have to make some settings? Or do I have to make some setting in frigate.yaml ?
Hi, thanks! First off, how are you running services on your Pi? Is it through Docker? Secondly, you'll need to deploy Frigate, configure it to use the Coral TPU, and then install the Home Assistant Frigate integration. This sounds like a lot, but it's pretty straightforward. The instructions in the video should work for a Pi.
@@Jims-Garage Hello ! Thanks for quick reply. I don't drive docks. What do you mean with you'll need to deploy Frigate, configure it to use the Coral TPU ? Where do i fing a good video for that ?
Can I run Frigate in a more rudimentary way? I am currently using MotionEye but it isn't being maintained very well any more and I imagine it's going to fall further behind as time goes on. Currently run it in a VM on XCP-NG and Coral doesn't really work well as a passthru. I am only really interested in viewing it and recording when basic motion occurs, don't need the AI aspects so much.
@@Jims-Garage Yeah ideally I would but I'm fully VM-based in my homelab so prefer to keep it virtual if I can. I'm going to see if I can set it up in a "basic" way for now. Basically just going to down my Motioneye container and try Frigate instead - fingers crossed! If it works real well I'll see if there's more I can do or need to do with the AI detection (my homelab is short on rackspace).
@@lewistaylor6695 yes, simply port forward (although there are some security steps you should follow like having a reverse proxy with SSL, I have videos)
Unify? Store in Cloud? Well, now I know you do not go to the top of my list… Modern cameras do have the detection included, so external AI/process/software are not necessary anymore.
Not sure I follow you. I use unifi cameras with frigate and store footage encrypted in the cloud. What is bad about that? I have complete control of all hardware and storage. The coral tpu means you can buy cheap cameras, lots of them, and have AI on them all. It's far more cost effective.
UDM-Pro = Bad naming for enterprise, grossly under-powered and very limited abbilities. That being said, what you are doing on Frigate is supported on UniFi Protect. For security, running VM's is for enthusiasts and those who enjoy complexity and dreams of compact shite. Run hardware and be done with it. The big one not mentioned, Frigate is very much Google entangled, no thanks.
In what way is frigate entangled with google? The license is MIT, which is fully open source and forkable. There are over 150 contributers, so if a company would close it, the community would very much keep it open. The image processing is done via Tensorflow and OpenCV. Yes, Tensorflow was originally developed by Google, but is also open source with the apache license. So again, F-up and the community will fork it. OpenCV was originally sponsored by Intel for that matter, but it leads an open source on its own. Tensorflow works best with a Coral TPU, yes. A very specific piece of hardware for Tensorflow specific calculations, but it is equally able to use a GPU at a slightly higher power per calculation. It's not entwined in that way either. It runs fully locally if you want to. So no google involvement there either. Android is very much google, but android alternatives like GrapheneOS and Cyanogen are very much not despite the framework being android. And there are a multitude of project available where the community has forked and re-opened a project. So in that regard, who cares who originally developed something. If the project is big enough, it will remain open. You say "just run hardware". But that is just basically it. You trade time for money. And granted UniFi appears to be one of the very few localized hardware solutions. But what does UniFi use on the background for image recognition? There is a high probability that it's Tensorflow :) But talking entanglement? You trade one company for another. At least frigate is open source.
@@patrickd9551 thanks, I agree with all of this. I totally understand that dedicated hardware is the ideal choice, however, udmp is not that hardware. No object detection, no external storage, and not even a redundant drive (or certainly when I had one).
Best to check the max size m.2 it will take Vs the m.2 coral tpu. Otherwise you could do a m.2 key adapter to a full PCIe slot. Failing that, USB coral.