Dev.SN ♥ developers
Sections
Dev.SN
siliconwafer writes "The US Department of Homeland Security (DHS) is looking to acquire a vehicle license plate tracking system, to be used at the national level. According to the solicitation obtained by the Washington Post, commercial readers, supplied by a private company, would scan the plate of vehicles and store them in a "National License Plate Recognition" (NLPR) database. This is already being done at the state level, and privacy advocates are up in arms, with EFF and ACLU suing California over their automatic plate readers. Now that this has potential to become a broad and national program."
[ED Note: "Shortly after the Washington Post broke the story on the national plate reading system, it appears the DHS has shelved their plans for the tracking system, by order of Homeland Security Secretary Jeh Johnson, at least in the interim."]
(Score: 2, Interesting) by hankwang on Sunday February 23 2014, @07:34AM
Can somebody tell me, a non-American, how automatic scanners would identify the state that the plate is registered in? From looking at license plate samples [wikipedia.org] it looks like the state is in much smaller print than the registration number. Moreover, they are absolutely not in any consistent format: at the top, at the bottom, on the side, italic font, straight font, or with very bad contrast [wikipedia.org].
Avantslash: Slashdot+SoylentNews for mobile [avantslash.org].
(Score: 4, Insightful) by Sir Finkus on Sunday February 23 2014, @09:21AM
I'd imagine they wouldn't be trying to OCR the state names, but would look at the overall design of the plate. With access to the state's database of valid plate numbers and which cars they correspond to, it'd probably be even easier. You could even make it even smarter and take into account cars that are in the area to shrink the pool further.
Additionally, it isn't particularly critical that the plate is scanned correctly each time if there are enough cameras. Eventually you'd end up tripping one.
(Score: 1) by hankwang on Sunday February 23 2014, @10:22AM
Huh, you mean, instead of OCR'ing the state name on the plate, you propose to do automatic recognition of the color, brand, and model of a car? That sounds like an order of magnitude harder.
Avantslash: Slashdot+SoylentNews for mobile [avantslash.org].
(Score: 1) by Sir Finkus on Sunday February 23 2014, @10:43AM
I suppose I didn't explain it well enough. The idea is that they look at the plate, then make an educated guess as to the state by looking at the colors on the plate. The could then cross reference that with the the registration database to verify. It should be fairly easy to determine things like color/size of the car. If the camera misses it, the next camera should have a decent shot to guessing correctly.
(Score: 4, Informative) by mhajicek on Sunday February 23 2014, @11:30AM
Each state has a unique combination of text color and background color on the plate.
(Score: 1) by Mesa Mike on Sunday February 23 2014, @12:40PM
Not just colors, but other design elements too. For example, the Texas plate has an outline of the state of Texas and a silhouette of a cowboy on a horse.
(Score: 2, Interesting) by hankwang on Sunday February 23 2014, @01:54PM
Unique? If you click a few pages on Wikipedia Category:Vehicle registration plates of the United States [wikipedia.org], that is not the impression that I get; for most states, there are several different color combinations depending on the year of issuing. Black or dark blue on white seems to be pretty common.
(Not a systematic search, just some random clicks).
Anyway, doing the recognition based on colors doesn't sound like a wise idea, since it should also work when it's dark (infrared flashlight). Recognizing small details might also be troublesome since there may be some motion blur.
Avantslash: Slashdot+SoylentNews for mobile [avantslash.org].
(Score: 2) by mhajicek on Sunday February 23 2014, @02:54PM
Thanks, I stand corrected.
(Score: 2) by Nerdfest on Sunday February 23 2014, @02:35PM
The readers are also generally tuned to bias towards the more common plates in the area (the ones I've dealt with anyway).
(Score: 1) by githaron on Sunday February 23 2014, @12:51PM
OCR would be one way. Another possibility would be to ignore the text all together and instead train a machine learning algorithm to classify pictures of license plates against a labelled data set. It would probably be more accurate at a distance than OCR would be for getting state information. If you are interested in that kind of stuff, Coursera has an interesting class on the subject: https://www.coursera.org/course/ml [coursera.org].