Raspberry Pi AI Camera takes inferencing load off the CPU Raspberry Pi has launched a camera module with AI smarts on board. But all that inferencing goodness comes at a price. Looking for all the world like a normal camera module – if ever so slightly thicker – the Raspberry Pi AI Camera is based on Sony's IMX500 Intelligent Vision Sensor, which combines a 12.3-megapixel CMOS image …
This is entirely down to the training data use to create the model. Lots of data on this on the Pi and Sony websites, but you would need to get a lot of photos of the different vehicles in varying weather and road conditions. 100s good, 1000s better. You might be able to use video then extract the frames to feed into the model. Processing the data will take a long time, 24hrs or so on a decent machine.
This is great. Love the subheading and will wait until they fix that b4 buying.
I want to build a CCTV that does face recog. So I know who strangers are. Luckily I live near a bistro whose patrons are somewhat noisy on occasion. With this linked to my intercom that can broadcast audio from the Pi, I can ID people and get the intercom to speak their name and all details on the net about them at a legal level of 90db.
Auto-focus is essential for this.
CCTV cameras have been able to differentiate large and small vehicles for some time, and they tend to use similar Sony camera modules.
There's almost certainly training data available that would help you set this system up, even if you have to do the final training yourself by pointing the camera at the road and showing it the size of a car, and a lorry.
A manual focus lens is always preferable for this kind of job, as focus won't wander as objects pass by.
On GitHub.We are evaluting a few of the models at the moment and they are well established. Not much between them really. Number plate recog is there, but it has a monthly cost as they pull the data from a subscription source.
Face recog and number-plate models plugged into your local Ollama (best under Linux if you can handle it - Mint) with RAG database of data you want added *see that top-notch article on here a little while back about setting up RAG with LLM.
P.S. A recent report found that more than 80pc of those working in AI/SI especially in Jail-breaking/Encoding hacks would be classified as mildly insane. HTH. This also affects LLMs. The smarter they get the more insane the might appear to the knuckle-draggers.
No votes
Obviously, it can't all be "built-in" - There must be a chunk of back-end look-ups going on once the camera has done it's initial - ummm - "triage"? Could all that back-end data sit on the PI itself, keeping everything self-contained, or would it absolutely require a connection to the wider world?
Or does it just depend on the complexity of what I'm trying to classify?
I have no use for the "AI" stuff. But I'm a huge fan of the Sony sensor. Fast, high resolution, reasonably low noise, and surprisingly uniform across the array. Slap that behind a decent C-mount lens and even a modest rPi 3 becomes a respectable digital camera. Slap it on a small reflector telescope and it breathes new life into that machine. Remove the IR filter, then point it skyward for more fun.
Fixed focus is not a problem when it's pointed at a bird feeder precisely 2m away :) Ive gotten some outstanding shots, and my kids have named all their favorite birds. Every my daughter away at college will occasionally SSH I to the pi cam to check up on "her" birds. Fun habit.
3k x 4k pixel for $70
This is the visual equivalent of the first 16bit audio ADC's. That's producing a signal on a 5V range where it changes by 78 micro volts (which needs a very stable reverence voltage) on a LSB change.
And that's now hobbyist grade.
So kind of impressive.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
