[Torgo]

>Alice installs app on smartphone. Alice puts smartphone in cradle on dashboard.

I tried this, but camera phones are (currently) too terrible. I would be curious if anybody more skilled than me (I'm just tinkering) could make it work.

[ChuckMcM]

Not all cameras in all phones are too terrible, that said having done the experiment, I found it is much easier to do license plate extraction from a video rather than from some sort of stills. Using an off the shelf "GoPro equivalent" on the dash of my car, post processing the video for plates was relatively straight forward with OCV[1] and now that I think of it, would make for an excellent tutorial on the technology. Were I to do it again I would provide a copy of GPS co-ordinates on the screen to make it a bit easier to tie the two together.

For me the interesting bit I was trying to figure out was this technique (car mounted) gave you license plate 'segments' which is to say the car in front of you may have been there for two blocks and then turned right giving you a segment that was two blocks long with a right turn at the end. If you had multiple subscribers, then you could match that segment up with a segment from another vehicle and then interpolate the path of any observed plate through an urban area. Since the general color and shape of the car is trackable, you can actually create a number of segments for different cars if you've captured their plate at any time, and if you manage 'untagged' segments (cars seen but you have yet to see a plate for) then you can add those as well. Needless to say, the taller vehicles are a "win" here for more complete surveillance.

[1] http://opencv.org/

[Torgo]

I was also using opencv, but I had a ton of trouble getting clean video from a couple different cellphones and a tablet (lot of blurring from movement) and was putting off buying a gopro to try. This is great info, thank you.

Some commercial units apparently have tail-light detection to narrow down scanning for plates.

[karambahh]

See Mappilary work on this area. Although they work on static pictures, I wouldn't be too surprised if their work could be succesfully applied to real-time video.

Another area you can explore are commercially available units. French police use on board license plates readers which seem to be quite fast (at least when you watch the french equivalent to the "cops" tv show).

http://www.teb-online.com/fr/analyse-images/lapi-lecture-aut...

LAPI(Automated Reading of License Plates) is the acronym you want to look for. I think several hardware and software manufacturers equipped different police units so if you're really interested, you'll be able to find several datasheets and derive a reasonable minimum computing power from them.

[burger_moon]

What if instead of using or in conjunction with using vehicle mounted cameras, you also utilized traffic cameras and street cameras. Many are accessible online.

[telmnstr]

Were you able to figure out the states?

[ChuckMcM]

No, that would have taken a lot more work. The algorithm was really primitive, find rectangles, look for letters and numbers which were > 50% the height of the rectangle.