An LLM isn’t going to help you here, but basic Computer Vision and a SQL database has been a solution _if you have the cameras_. I wrote a license plate reader as a university project using OpenCV almost 20 years ago.
One of the risks of LLMs is that a lot of tasks go from "an expert could do this easily given a few weeks" to "anyone who thinks to ask an LLM can do this easily and get results the same day"
By that logic, all problems are solved with LLMs, though.
My friend, if you have a database of license plates extracted from single images taken by multiple cameras, YOU ARE TRACKING UNIQUE VEHICLES ACROSS A REGION.
Terabytes of data don't matter because you don't need to search terabytes, you need to search a few MB of text data. You don't even have to store the original video.
Amazing how you can move the goalposts to make things impossible, isn't it? Where in the world did "without compute" come from? Are they not even allowed a decent desktop computer?
I did this with Gemini 3, mostly for fun and to test it's capabilities. Teslausb records all dash cam videos and auto syncs it to my nas when in wifi range. Yolo and opencv extracts and does ocr on any defected license plate, and puts it all on a map, along with trip information. Not particularly useful or interesting, and not something I would have done pre-llms, but the difficulty was basically writing a one paragraph prompt and using some free tokens
There's a very important difference between "anyone could walk through my door and steal my stuff" and "this person walked in my door and stole my stuff".
Flock cameras are roughly that secure.
"I gave the person keys to my house and then I trusted they wouldn't open bathroom doors while somebody was there".
Like law enforcement is being given access to the systems, the door isn't "left open", a key was given to them.
Here's another way to look at this. Municipalities are the primary operators of ALPR cameras. Any municipality that would scan bumper stickers looking for Trump opponents is not going to be receptive to any appeals for regulation.
One problem with this whole debate is that people are coming to it with movie plot concerns rather than understanding what's actually happening with them. That wouldn't be a big deal if this was a slam dunk public policy case, but it isn't: there is broad bipartisan support for these devices.
There are deeply problematic things happening just with license plate pings!
They're matching specific descriptions of cars to incidents, like, "this vehicle has been present at the site of 5 previous package thefts".
- Who is "They"?
- Why do you say "nobody has time for that"? What is "that"?
- Why are you dismissing genuine concerns through unhelpful language like, "coming to it with movie plot concerns".
- Why wouldn't "that" be a big deal? What is "that"?!
- What are the deeply problematic things?
- "They're matching specific descriptions of cars to incidents" -- no they're not. Just looking at Bloomingdale's audit logs, there are 13k examples of searches done for the simple reason, "suspicious".
- Why does municipalities being the primary operators matter?
Asking from a place of genuine confusion by how you think about these things.
Police setting up a 1984 monitoring system throughout your city, tracking every car, person, activity -- yields lots of questions, oversight, concerns, debate, challenges, etc.
Some private business doing the same, and then letting the same police use it at will as a paying customer -- yay, all of the invasive monitoring with none of the oversight.
Privacy laws now.
I'm frequent surprised by how many people think that privacy laws block the police from recording their activities in public. For whatever reason, Flock is getting a lot of press, but this is hardly a new field.
> However, ANPR did not become widely used until new developments in cheaper and easier to use software were pioneered during the 1990s. The collection of ANPR data for future use (i.e., in solving then-unidentified crimes) was documented in the early 2000s.
- Benito Mussolini
Is a bit misleading itself, to do this at scale requires all those iffy data centers.