I recently set up Home Assistant OS on a Raspberry Pi 5 (8gb) for my parents to do object detection and 24/7 recording via the Frigate NVR add-on, but the detection inference speed leaves a lot to be desired. On my personal set up, I have been using the USB Coral TPU on an Unraid server box and it works incredibly well; however I have heard mixed results when it comes to using one with a RPi due to the limited power delivery via the USB controller; not to mention that it is relatively difficult to find and $60 more expensive than the M.2 variant. I read many posts across the web demonstrating success with using the M.2 Coral on a bare metal RPi running debian, but cannot find much in the way of success stories for installing one via Home Assistant OS, specifically on a Pi 5. Can anyone confirm that they have been able to do this successfully before I make the plunge and purchase one in addition to the required PCIe breakout HAT? If so; was it more or less plug and play, or are there some manual installation steps related to getting the driver to play nice with it? Any insight would be greatly appreciated.
There are some cheap x86 PCs that are better than Pi5 at same or similar cost
Even used hardware is reliable and cost friendly
I am aware but I am trying to work with what I already have.
I currently have a working instants of frigate running in a docker on a raspberry pi 5 8gb using a hat that has dual m2 slots and a PCIe coral TPU.
I’m pretty sure in order to get it working I had to install it using the desktop version of the OS (not server)
In future I’m planning on moving my home assistant instants to the same raspberry pi.
For more detail I’m getting a detection inference speed of about ~7ms for 4 camera feeds at 720p at 5fps.
I built the same setup with Pi 5 and a coral dual edge TPU. I have 10 camera streams. I am using the 27W power supply and active fan coolers. The setup works. Frigate shows CPU 100% used. But the CPU temperature is 47°C
However after few minutes entire Pi5 freezes.
Did you have similar issue?
Any idea what I maybe doing wrong?
Thanks!
Don’t expect wonders with using a Coral TPU. 10 camera feeds is alot of data the CPU has to process. The Coral TPU does little more than object detection while the CPU still has a big amount of data to process for decoding the video streams. This also depends on how you configured the resolution and the frame rate of the streams. At this point hardware acceleration comes into play. Unfortunately I have no clue whether the Pi5 CPU supports hwaccel args.
I suggest you read about the recommended hardware for Frigate HERE. Especially the paragraph “What does Frigate use the CPU for and what does it use a detector for?”. It will help you to undestand what the coral dual edge TPU can do and what it cant do.