• Advertisement
Sign in to follow this  

car plate reader

This topic is 4477 days old which is more than the 365 day threshold we allow for new replies. Please post a new topic.

If you intended to correct an error in the post then please contact us.

Recommended Posts

Advertisement
Heh. Well.

Now, in the general case--a photograph, sometimes grainy, sometimes blurry, of zero or more cars in arbitrary positions with less-than-spotless license plates--this is a VERY difficult problem. Not as difficult as many other computer vision problems, but it's up there. Reading an unoccluded license plate from a normalized, axis-aligned region is much easier. So here are some questions for you:

(1) exactly what conditions does it have to put up with?
(2) what is your experience with computer vision, including topics such as OCR, edge transforms, etc?
(3) how much time do you have?

Share this post


Link to post
Share on other sites
Humm.. check some websites regarding convolution filters. I was looking into these a few months ago and found a website talking about using them for exactly this.

Share this post


Link to post
Share on other sites
Quote:
Original post by Sneftel
Heh. Well.

Now, in the general case--a photograph, sometimes grainy, sometimes blurry, of zero or more cars in arbitrary positions with less-than-spotless license plates--this is a VERY difficult problem. Not as difficult as many other computer vision problems, but it's up there. Reading an unoccluded license plate from a normalized, axis-aligned region is much easier. So here are some questions for you:

(1) exactly what conditions does it have to put up with?
(2) what is your experience with computer vision, including topics such as OCR, edge transforms, etc?
(3) how much time do you have?





let's see,
first of all, i have a picture as a bmp, this picture is taken every time a car approaches towards a camera...
as a result, the plate of the car is always straight on the x and y axis, it only may be rotated on the Z-axis, so instead of having the plate as a perfect rectangle, it will be with a little perspective....


i was thinking to do the following:
- i have to start by finding where the plate of the car is on this bmp, and create a new bmp containing only the plate
- apply a formula that makes me transform the perspective rectangle to a perfect rectangle, so it will be axis alligned and i can normalize it to a specific size, since i have the dimensions of the plate rectangle in the pic once it is extracted from the general image
-read the characters(only numbers).


the first problem i have, is how will i find the boundaries of the plate? the plate is a lebanese car plate, it is a rectangle with black borders, white from inside and contains the plate numbers...
this is the problem i am facing...

how am i going to find the region of the plate so i can extract it!


are the steps i am thinking of are correct?
does someone have another solution, like directly reading the numbers from the image?

i only havea week for the whole project!

Share this post


Link to post
Share on other sites
They gave you too little time to wok on this project. License plate recognition is pretty complex task. Here you can find a diagram of how such program may work. They even have Matlab implementation. It's about 100 kb of code, so mere rewrite to c/c++ may take longer than a week. And then you will probably need to change it a little since they use yellow LPs.
HTH and good luck [smile].

Share this post


Link to post
Share on other sites
Quote:
Original post by b2b3
And then you will probably need to change it a little since they use yellow LPs.

He will probably have to change a lot more than just a little. Yellow LPs have the advantage of a larger colour contrast to the immediate image environment. Usually, you don't have any other yellow parts on the front part of a car (except for cars painted yellow). In that algorithm, they take advantage of the colour contrast to identify position and perspective transformation of the license plate. This will also work more or less well under suboptimal lighting conditions (twilight, night, fog, etc) because yellow is a bright marker colour, with a limited reflective wavelength.

The OP said that lebanese LPs are white with black border. Good luck with that. White has no colour contrast, ie. no limited wavelength. White will also radically change its appearance under varying lighting conditions. Also, a lot of cars have white or whiteish parts on their front, so that will make detection harder. You might be able to use the black border as a marker, but this depends on how the LP actually looks like. Can you post a pic of a lebanese LP ? I've never seen one.

All in all, doing this reliably is a very difficult problem. Doing this in one week by a single programmer is impossible. Unless you already have extensive experience with computer vision, and access to appropriate libraries and tested code. And even then, one week would be a rather insane deadline. The learning and training phase of the neural networks alone would take several weeks, with hundreds of tests under varying lighting and environmental conditions.

My advice: if this is a school project, then tell them to change the subject. If this is a paid contract you accepted, well, then you're in trouble.

Share this post


Link to post
Share on other sites
Sign in to follow this  

  • Advertisement