Re: ANPR/Car number-plate recognition

Re: ANPR/Car number-plate recognition

Posted by: "beat.naef"
beat.naef@yahoo.com
 

beat.naef

Mon Apr 13, 2009 3:11 am (PDT)

Monica,

I uploaded a small image set taken by an IR LPC to the file/LPR Project dir and some BlobAnalysis-based code.



Are you working on a LPR project also? If so would you mind sharing your experience and code?



Regards

Beat

--- In OpenCV@yahoogroups.com, Monica Guzman <monait2003@...> wrote:

> Hi, can you send me a lot of images and videos obtained from your ir camera....

> Grettings..


> --- El vie 10-abr-09, beat.naef <beat.naef@...> escribió:



> De:: beat.naef <beat.naef@...>

> Asunto: [OpenCV] Re: ANPR/Car number-plate recognition

> A: OpenCV@yahoogroups.com

> Fecha: viernes 10 de abril de 2009, 10:26



> I have also started a LPR project using OpenCV. I am using
gray-scale images taken by a true IR LPC (Bosch/ ExtremeCCTV Reg-X,
combined with Axis 241S video server to capture jpg image sequences
(M-JPEG)) that is installed in the field (actually several LPC
installations) .

>

> I have tried a number of approaches:

> 1. using the square.c code in the OpenCV src samples (experiment
with different image smoothing, morphing techniques, etch/ contour
detection techniques)

> 2. tested with some SIFT code

> 3. in the process of testing haar-like feature method

> 4. in the process of testing background averaging method

>

> Approach 1 was initially promising. But when I tested about 100
randomly selected images against it, the success rate was not very high
(that maybe a reflection of my lack of experience in computer vision
technologies and to fine-tune object recognition) .

> Approach 2 initial tests were not very exciting because not too
many matching candidates points were found. However, I have not spend a
lot of time to fine-tune and experiment

> Approach 3 still need to adapt code and learn more about the technology

> Approach 4 same as 3

>

> I am interested in sharing my experience and data sets and
starting an OpenSource project on SF anybody interested? Anybody
interested?

>

> Beat

>

> --- In OpenCV@yahoogroups. com, Alper Yaman <alperyaman@ ...> wrote:

> >

> > You can use bottomhat transform to get rid of trouble in case
of the same  color of car and plate. In addition, you can use radon
transform to correct skewing. I tried so many methods to divide
characters from the plate candidate region. First I extract the exact
plate rectangle from the plate candidate region. Then I apply global
threshold and labeling to get characters as objects. I elect the
noncharacter object by some rules (area, width/length etc.). This
method has some trouble while the exact plate region is somehow
connected to the other regions. Does anyone have suggestion?

> >

> > Alper

> >

> >

> > --- On Sat, 1/24/09, Carlos Frederico Mendes <opencvlist@ ...> wrote:

> >

> > From: Carlos Frederico Mendes <opencvlist@ ...>

> > Subject: Re: [OpenCV] ANPR/Car number-plate recognition

> > To: OpenCV@yahoogroups. com

> > Date: Saturday, January 24, 2009, 3:36 PM



> > Hi,

> >

> >

> >

> > I'm a novice in this area, but i already worked on a car license plate

> >

> > recognition program, i think you will have trouble to detect the rectangles

> >

> > in case the car have the same colour as the background of the license plate.

> > Fred.


> > On Fri, Jan 23, 2009 at 7:30 AM, ckd600 <alanp23@gmail. com> wrote:


> > > Hi all,


> > > Whilst I await delivery of my Learning OpenCV book, I'm thinking of

> >

> > > projects I'd like to look at. One is real-time car

> >

> > > license/registratio n plate alpha-numeric character recognition.


> > > It's probably worth mentioning I had this idea in mind for UK plates

> >

> > > which are standard size, font, and colour.


> > > My first idea of how it would work was like this:

> > > - Detect rectangles(potentia l number plates) in an image

> >

> > > - Extract each rectangle found and try to split in to characters (not

> >

> > > sure how I'm going to do this, maybe I can detect the spacing between

> > > the characters)

> > > - Run each split character through a SURF scan of the standard UK font

> > > character set and choose the highest % match

> > > - Move to the next char



> > > The only problem I can see is that it wont match if the image is

> >

> > > skewed, but that can be overcome by carefully planning the camera

> >

> > > position. An advantage is that it wont need to be trained, it will

> >

> > > just need a character set to work from.



> > > What do you think?

> >