Hello Dustymack,
The general approach apparently taken by several vendors can be
summarized as follows:
- capture image(s) and select the best (or several good) frame(s)
- select the "license plate" portion of image / rectify
- isolate the characters for recognition
- identify the characters
followed by some post processing steps where the identified characters
are used (usually a database look up). For an example of this, see
http://visl.technion.ac.il/projects/2003w24/
and scroll down a little for a flow chart and several samples of the
steps used in this student project.
http://www.licenseplaterecognition.com/
A tutorial on license plate recognition by one of the vendors.
Includes how license plate recognition can be used for a variety of
applications.
Specific methods used include:
- template matching; match each letter to a specific example. Noted
in several papers that this can be quite difficult to apply in
general.
- structural matching; you look at the shape and connections of a
letter. For example, comparing between B, D, 9, and 6 you could
recognize the loops and relative position (top, bottom, center) and
size but have difficulty distinguishing betwen B and 8.
- neural networks; trained with several samples can be resistant to
noise but may require a large investment in time / training with
different character shapes (font).
A very good summary of the issues with automatic recognition of license plates at
http://www.photocop.com/recognition.htm
Scroll to the bottom of this reference as well for an extensive list
of suppliers. If you are interested further in the use of photography
in trafic enforcement, I suggest starting at the home page
http://www.photocop.com/
and checking several other links as well. Some describe methods used
to obscure or work around the use of photo systems in police
enforcement. There is also another list of companies (with links) that
provide more turn key solutions.
A Dutch site, using video capture plus speed measurement to help
enforce speed limits and fining the car owner.
http://itctraffic.com/videoenforcement.htm#License%20Plate%20recognition
Scroll up / down for several good sections of information. In
particular note the claimed system accuracy of over 99% on
"identification" (apparently is a vehicle present) but only 78% on
license plates. Also note this takes images at three locations (start,
800 meter, 3000 meters) for the speed measurement so there is a need
in this system to recognize the "same vehicle" in each image.
Apparently the Dutch also found it necessary to change the font of
their license plates to improve recognition. See
http://www.sunpig.com/martin/archives/2003/09/20/dutch_car_license_plates_and_traffic_control/
for an example.
Somewhat short on details but does talk about resolution, types of
cameras, and related issues at
http://www.machinevisiononline.org/public/articles/archivedetails.cfm?id=1206
A look at the more general problem of identifying text in a video and
extracting that text:
http://www.informedia.cs.cmu.edu/documents/vocr_ieee98.pdf
Also has references to license plate capture from video (1997 IEEE
conference paper).
A recent conference (October 2004) with several OCR / vehicle license
plate recognition papers
http://www.ieeesmc2004.tudelft.nl/?menu=program.&slotid=149
A couple hundred references to techincal papers found with the search phrase
license plate recognition citeseer
For example, a couple clicks from the second reference brought me to
http://citeseer.ist.psu.edu/572231.html
A paper titled "New Methods for Automatic Reading of VLP's (Vehicle
License Plates)" which includes several references to / from this
paper as well. I strongly suggest using citeseer for technical
references if you have specific concepts or phrases to research.
Google also has a recent service at
http://scholar.google.com/
where I entered the phrase
vehicle license plate recognition
and had another 800+ references. Note that in this case, the
references are free but the full document may require a fee for
access.
A FEW vendor sites:
http://www.anpr.net/041129/index.htm
http://www.platerecognition.info/1102.htm
Adaptive Recognition Hungary, producer of both hardware and software
for license plate recognition.
http://www.saic.com/products/transportation/iis/
SAIC provides turnkey and custom systems for recognition as well as
integration with with related systems (e.g., recognize loads on
trucks, driver photo identification). SAIC claims over 95% accuracy on
identification numbers with higher rates if backup data is available.
Also indicates a rate of vehicles that can flow through their
recognition site (roughly 100 to 300).
http://www.cormactech.com/neunet/download.html
Not specific to license plates, but this vendor of Neural Network
software includes an OCR data sample to train / recognize if you want
to get a feel for the issues involved.
http://www.htsol.com/Products/SeeCar.html
High Tech Solutions, provides several related vehicle license plate
systems. Claims use at 2004 Olympics in Greece.
http://www.jrwald.com/blue/license/newtech/LPD.html
A more restricted system, recognizing license plate numbers during
manufacture of the plates (but very high speed - 1000 per hour).
Free and for fee software / algorithms are available from links at
http://www-2.cs.cmu.edu/~cil/v-source.html
for a variety of computer vision applications (including license plates).
Search phrases used include:
ocr automobile license plate
automatic recognition automobile license plate
license plate recognition
--Maniac |
Clarification of Answer by
maniac-ga
on
23 Jan 2005 18:13 PST
Hello Dustymack,
The sites that describe specific methods and accuracy appear to be
even more rare than I expected. The following summarizes what I could
find for specific systems.
[1] IPMS Vehicle License Plate Recognition System
http://www.singaporegateway.com/optasia/techinfo.htm
Methods used for image capture:
- standard CCTV, using a single frame (even or odd) sized so
characters are at least 18 pixels high
- lighting, options for continuous or strobed, visible or near IR
- single or multiple images captured
Methods for OCR:
- breadth first AI for locating plate location in image
- no real specifics on OCR method except it allows for "fuzzy"
matches, most likely a neural network with confidence levels
Sample plates read / accuracy:
http://www.singaporegateway.com/optasia/plates.htm
This vendor claims over 99.7% accuracy for Singapore plates.
[2] US Customs Example
http://www.cbp.gov/xp/CustomsToday/2001/December/custoday_lpr.xml
Methods used for image capture:
- high speed "video camera"; says sensor is sensitive enough for
1/10000 second image but does not state it captures at that rate
- lighting appears to be near visibile IR strobe
- image capture of both front and rear of vehicle
Methods for OCR:
- text implies it categorizes plates by issuing agency (e.g., US
state, Canadian province, presumably to aid in recognition
- no detail on method for individual character recognition
- has back end interface to national databases for verification; may
be used to "correct" recognition of plates
Sample plates read / accuracy:
http://www.singaporegateway.com/optasia/plates.htm
Article claims 90% accuracy for Canadian, Mexican, and US plates.
The other online references I found would either describe methods or
accuracy but not both in the same context.
Another reference I happened to find while searching that you may find interesting.
http://users.erols.com/lnelson/lpir.html
Shows a series of images captured in natural light and comparing to
use of near IR lighting and cameras (scroll down to section starting
with "Optimizing License Plate Design..."
--Maniac
|
