I am new to the Raspberry Pi world, and it seems through long research on Gemini and coding trials through Cursor, It’s a deep rabbit hole.
I have already built an Edge AI fatigue monitoring system (FRMS) as a multilevel UI, including a very complex risk processing backend. My day job is as a safety/risk professional, so the subject is core business for me.
What isn’t core business, is assembling the hardware for the in-vehicle part of my fatigue monitoring system. What I was seeking was some kind of feedback on my choice of hardware so far, before I make the purchase commitments to make sure it will do what I need - live face and eye position calculation through YOLO or whatever works for that, fed through the in-vehicle AI, then lightweight JSON back to the server/db for interaction with the risk engine. I have the whole SW system designed and working, apart from the camera system. This may sound similar to other enterprise systems in the market now, but this is vastly different in implementation and outside the scope of this post.
Here is my hardware list with some rationale for the choices (as explained to me by Gemini/Cursor or actual people):
Raspberry Pi 5 4GB - chosen for price and availability, and because the AI HAT does the heavy lifting with this concept
Pi M.2 HAILO 13 TOPS HAT - looking for 30 fps at medium resolution (the cameras don’t run concurrently)
USB 3.0 NVMe PCIe GEN 3 storage with 128GB SSD as the boot drive (USB 3 channel 1) - to get around the microSD card bottleneck, and price.
USB 3 powered hub (USB 3 channel 2) - so not to draw excess power from the Pi to run the dual cameras
2 x Arducam B0332 cameras running on the USB channel 2 - chosen because they are NoIR filter
940nm LED arrays power by the 12v supply
Pi wall power brick for the bench dev
12v wall brick to power the USB hub
Will this do what i’m trying to achieve do you think?
Hey @Rod75259, that hardware list looks sensible for a prototype. A couple of quick notes:
I’d include an active cooler, as the Pi 5 can get quite warm under sustained camera/AI workloads.
Typically when people use cameras with the Pi they use the onboard CSI connectors, and if you’re using the Pi 5 you have two to choose from. I would probably use something like the Raspberry Pi Camera Module 3 (SKU: CE09425) instead, or the Raspberry Pi Camera Module 3 NoIR (SKU: CE09427) if you need.
Thanks for the response. I was hoping for some helpful feedback and looks like I got some
I have the active cooler.
As I understand it, these are the issues:
The environment this unit is going into is not friendly (commercial vehicles)
My reading shows that the CSI connected cameras are not robust enough for the above env.
I also need to power 940nm LED’s for the cameras, so the USB hub was in place to do that (using heavy gauge cable to minimise resistance). One cable being around 1.5m, the other maybe 2m.
The cameras must be monochrome to provide a simpler native format for the Pi unit to process.
Edit: further research shows the cameras you mentioned are rolling shutter - this won’t work as they need to be global shutter cameras to work with YOLO etc.
What are the options for a robust cable and power setup using the cameras you suggested?
Fair enough then. There is an official Raspberry Pi Global Shutter Camera which would be the better quality camera, but it uses the CSI lane so that’s out. Even if you get CSI cables that are long enough, USB is going to be better.
In the vehicle, you are going to need something like the PD Power Extension Board for Raspberry Pi 5 (SKU: TPH-EP-0225) to run the Pi 5. The Pi 5 needs PD Delivery to power it, so no step down converters unfortunately. You can mount it beneath the Pi 5 so it doesn’t interfere with the rest of the project.
Actually, while we are here, there is something I did not notice in my first post. If you’re going to have both an NVMe HAT and an AI HAT you are going to need a PCIe Expansion Board to turn the Pi’s one connector into two: PCIe Expansion Board For Raspberry Pi 5 (SKU: WS-30490)..
I know that’s quite the stack so you can probably mount the PD Delivery separately or work with a different PD Delivery system with 25W, but at a minimum you will need the PCIe expansion board for your setup to work.
In regards to “If you’re going to have both an NVMe HAT and an AI HAT”, I dont think i’ll be doing that. This is the reason for the USB 3.0 SSD, im only using the AI HAT. Knowing my setup, would I still need the PCIe expansion board?
Now you have raised the issue of power, the PD board looks like the ideal solution I didnt know existed. I can power the PD board using a typical 12/24v buck converter?? I guess that means the power then going to the Pi is what it expects.
Ahk, never mind, I just reread your post incorrectly and thought the SSD was going to be using the PCIe lane as well. With your setup you won’t need the PCIe Expansion Board.
The PD Board is great. You don’t even need the buck converter. If you power it from the DC Jack you can accept anywhere from 7 - 24V input. Then just connect a USB-C cable from the PD port to the Pi 5, and the Pi will receive its 5V 5A, exactly what you need.
