Amazon's Mechanical Turk (https://www.mturk.com/mturk/welcome) can be used as a cost-effective way to get human operators to tag license plate images for training data.

For this to work we would need a simplified, web-based training workflow. It would need:

• An introductory training where users could be trained on the type of data needed
• A validation set so that users who do not enter correct data would be disqualified
• Training workflow for both plate regions (detection) and plate characters (OCR)

OpenCL support can be added for license plate detection, binarization, and (in the next version of Tesseract) OCR. This would significantly improve the recognition speed..

This command:
./alpr /home/mhill/webservice/temp_uploads/5fa0beb7-002f-4b68-a59d-042c5ab18757.jpg -dj

on this image:

Produces this output:
{"epoch_time":1409713153536,"processing_time_ms":303.881430,"results":[{"plate":"DRUPAL","confidence":95.113091,"matches_template":0,"region":"��#���","region_confidence":1,"processing_time_ms":238.184692,"coordinates":[{"x":187,"y":134},{"x":292,"y":134},{"x":292,"y":182},{"x":187,"y":182}],"candidates":[{"plate":"DRUPAL","confidence":95.113091,"matches_template":0},{"plate":"0RUPAL","confidence":85.834694,"matches_template":0},{"plate":"ORUPAL","confidence":85.360779,"matches_template":0},{"plate":"BRUPAL","confidence":84.456360,"matches_template":0},{"plate":"DR0PAL","confidence":83.963173,"matches_template":0},{"plate":"QRUPAL","confidence":83.164261,"matches_template":0},{"plate":"0R0PAL","confidence":74.684776,"matches_template":0},{"plate":"OR0PAL","confidence":74.210861,"matches_template":0},{"plate":"BR0PAL","confidence":73.306442,"matches_template":0},{"plate":"QR0PAL","confidence":72.014343,"matches_template":0}]}]}

It would be great if there could be an option to ignore ie don't output a continuous stream of plate results and associated images if openalpr sees the same plate for multiple frames.

Hello,

I used train-ocr to make it learn the switzerland fonts and added it to the eu dataset. When doing some tests with ocr debug set to 1 I was surprised to see that the font used was still the "germany" one, despite analyzing a swiss plate. The result for the plate is correct, tho.

The code showing the font is:

const char* fontName = ri->WordFontAttributes(&dontcare, &dontcare, &dontcare, &dontcare, &dontcare, &dontcare, &pointsize, &fontindex);

Do you have a short explanation of how tesseract is used and how it choses which trained font to use? does it try all of them and pick the best match?

Just in case there is a crash -- this would make it much easier to identify the bad code:
http://stackoverflow.com/questions/77005/how-to-generate-a-stacktrace-when-my-gcc-c-app-crashes

h264 is more network efficient than mjpeg. Add support in the alpr daemon as a configurable stream option

Smaller characters (lower resolution plates) should have lower confidence. When the plate gets closer and the characters are large, the confidence should reflect the higher probability of an accurate read. Since plate image sizes are normalized, the cropped size has to be mapped back to the original image to get an accurate size.

The example below had to fall back to a best-guess for plate edge. The area identified by plate detection was too small. Given that the characters were properly identified, the plate area should be recropped (only in cases where it is too small based on the character height) before sending to platelines. The plate edges are strong and would likely be properly identified.

From benchmark db: us/car9-1.jpg

If we could have an option to save the clipped images too that would be good, and helpful to automatically gather appropriate images for training OCR.

Steve

Hello,

The doc says:

template_max_width_px = [maximum width of the plate before processing. Should be proportional to the plate dimensions]
template_max_height_px = [maximum height of the plate before processing. Should be proportional to the plate dimensions]
min_plate_size_width_px = [Minimum size of a plate region to consider it valid.]
min_plate_size_height_px = [Minimum size of a plate region to consider it valid.]

And then when I read the code I see that:

  Size minSize(config->minPlateSizeWidthPx * this->scale_factor, config->minPlateSizeHeightPx * this->scale_factor);
  Size maxSize(w * config->maxPlateWidthPercent * this->scale_factor, h * config->maxPlateHeightPercent * this->scale_factor);

Which actually correspond to:

max_plate_width_percent = 100
max_plate_height_percent = 100

This is rather confusing and unclear. I propose a renaming of the variable and some cleanup, maybe remove max_plate_width_percent entirely and use template_max_width_px which would be renamed to max_plate_size_width?

I started a pull request and realised I wasn't sure of what I was doing, can you help me go in the correct direction? Right now it looks like you specify min plate size in terms of pixels, max plate size in terms of %, and max plate size in terms of pixel but just for ocr?

Anyone has ideas about his? I did build tesseract 3.02 in VS2012 by myself. When I build alpr, quite a few link errors like below happen. Thanks for your help.

Error 732 error LNK2019: unresolved external symbol "public: float thiscall tesseract::ChoiceIterator::Confidence(void)const " (?Confidence@ChoiceIterator@tesseract@@QBEMXZ) referenced in function "public: void thiscall OCR::performOCR(class std::vector<class cv::Mat,class std::allocator >,class std::vector<class cv::Rect,class std::allocator<class cv::Rect > >)" (?performOCR@OCR@@QAEXV?$vector@VMat@cv@@v?$allocator@VMat@cv@@@std@@@std@@v?$vector@V?$Rect@h@cv@@v?$allocator@V?$Rect@h@cv@@@std@@@3@@z) I:\WorkingProjects\References\openalpr-build\openalpr\ocr.obj openalpr

Europe, Japan, Pakistan all use 2-line license plates. Supporting these plates will require training data and modifications to the Character Analysis code.

Hello,

When using -j, it only prints the "best match" and not all the matches like -n does on the command line.

Is this intentional? it's kinda annoying because right now my plate is often in the top 5 results but not always the top 1, and parsing the output without -j kinda defeats the point :)

Using sparse mode optical flow tracking, the library should be able to track plate regions from frame to frame in a video. This would improve accuracy by:

• Reducing false positives. A plate must be seen multiple times to be considered valid
• Providing more reads per plate (higher accuracy reads would be used)
• Remove double-counting of reads

This will likely be very processor intensive, and should be implemented as a GPU-only feature.

I am using Visual Studio 2013 to try to build OpenALPR. After fixing numerous errors I have reached a point where I can go no further. I don't know if I am using a newer version of tesseract, or if something else is going on, but here is my problem:
When I build Tesseract I cannot find a tesseract.lib file (Which is needed for classifychars and benchmark projects). I tried replacing tesseract.lib with libtesseract302.lib in the classifychars property pages under linker->input->additional dependencies, but then I get compiler error "LNK2019: Unresolved external symbol public:float __thiscall tesseract::choiceIterator::confidence " when I try to build classifychars. (Note that there are total of 6 errors when I build, 5 are similar and they all have to do with the choiceIterator: constructor, destructor, Next, GetUTF8Text, and of course confidence. The last is error LNK1120, which is a count of the other errors.) I am using Tesseract 3.02.02. I am new to Tesseract, OpenCV, and OpenALPR so any help would be greatly appreciated.

Hi, I'm trying to install using the easiest method on README.md
after I run the command sudo apt-get update I get this error:

W: Failed to fetch gzip:/var/lib/apt/lists/partial/deb.openalpr.com_master_dists_openalpr_main_binary-amd64_Packages Encountered a section with no Package: header

E: Some index files failed to download. They have been ignored, or old ones used instead.

what can I do to fix it?

update: when I use the easy method with contains downloading dependencies and compiling openalpr I get this error:
-- Found Tesseract
CMake Error at CMakeLists.txt:47 (MESSAGE):
OpenCV version is not compatible : 2.4.1

-- Configuring incomplete, errors occurred!

See:

https://github.com/openalpr/openalpr/blob/master/src/openalpr/config.cpp#L164

(Lines 163-165):

    float stateIdImagePercent = getFloat("common", "state_id_img_size_percent", 100);
    stateIdImageWidthPx = round(((float)templateWidthPx) * ocrImagePercent);
    stateIdimageHeightPx = round(((float)templateHeightPx) * ocrImagePercent);

Perhaps the last two lines should be using stateIdImagePercent?

As it says on the tin. From my point of view, if the jpg name could be top plate result, date and time and % confidence that would be much better than UUID. Some options would be great.

Cheers

Steve

Could you tell me how to configuration the Tesseract_DIR ?

In the README.we can see blow:

• SET(Tesseract_DIR "/home/mhill/projects/alpr/libraries/tesseract-ocr").

I was down tesseract-3.00,install to my ubuntu and test was ok.But when i make openalpr.It tell me [ocr.h:37:21: baseapi.h：There is no this file].

I think i configuration the Tesseract_DIR error. now was [SET(Tesseract_DIR "/git/tesseract-3.00/api/.libs")] , And Tesseract is in /git/tesseract-3.00/。

tesseract path

Error message

Alprd can rotate input image by x degrees to get better accuracy. If the camera is fixed and rotated x degree, alprd can compensate for this by rotating the image before analyzing

I find that include guards like "ALPR_H" and "OCR_H" are too short for the safe reuse of your header files.
How do you think about to make them unique by appending a kind of UUID?

Hi,

Good work!

I am testing the library as a plugin for the ZoneMinder CCTV software.

The alpr object is loaded dynamically, plates are detected but every time I delete the alpr object I get a segmentation fault.

I found the cause: Conflict of class name

Class "Detector" and "Config" are defined in OpenALPR but also in ZoneMinder

Although these classes are used only internally in the OpenALPR library, the duplication of class name leads to unattended errors.

I temporary fixed the issue by enclosing the conflicting OpenALPR classes in the namespace "alpr".

To avoid this kind of problems (especially with very common class names), the OpenALPR library should provide his own namespace.

Emmanuel Papin

The README.md claims "OpenCV v2.4.x " is required for compiling. Build scripts indicate OpenCV >= 2.4.8 is required.

From ../src/CMakeLists.txt about line 55:

IF (${OpenCV_VERSION} VERSION_LESS 2.4.8)
MESSAGE(FATAL_ERROR "OpenCV version is not compatible : ${OpenCV_VERSION}")
ENDIF()

How to build classifychars utility and other from misc_utility.
Automatically they not building binary.

Hello,

I'd like to use openalpr to scan various types of license plates.

1. Is there an easy way to create a new plate format?
2. Is there an easy way to have a "catch all" format which yields good results?

In general, how do you handle license plate recognition when you expect several formats?

[EDIT] I just found about the -d option, is that the way to go?

Some countries use characters that have characters which are not fully connected (e.g., Serbian zero characters). The height filter (in character segmentation) should be configurable per-country and should be turned off for EU.

Copied from Roya Gh forum post:

I want to add support for plates of a new country to OpenAlpr so that it can recognize the license plates of Iran.
I've tained tesseract for persian and it works for Iimages of Persian text,
but when I try to use OpenAlpr the output is sth like this:

plate0: 9 results
- лллллилл confidence: 83.9618 template_match: 0
- лллллил confidence: 81.1455 template_match: 0
- ллйлллил confidence: 81.0826 template_match: 0

I've tarined for unicode but I've realized that OpenAlpr doesn't use UTF8, I suppose that the problem is due to this,

const char* choice = ci.GetUTF8Text();
postProcessor->addLetter(*choice, j, ci.Confidence());

here https://github.com/openalpr/openalpr/blob/master/src/openalpr/ocr.cpp#L104

Is there any implementation of OpenAlpr supporting UTF8?

There are certain areas in the input stream that we don't wish to attempt detection (e.g., sky, a fence area, a timestamp). Add a black/white mask image as input (resolution must match input stream) that will not attempt detection.

This should be implemented in such a way that it both decreases false positives and improves performance (since less of the image must be analyzed).

The alpr binary should report that no plate is found when recognition is unsuccessful. Json output should remain as-is (i.e., returns an empty set: [])

Hi,

I've noticed the letter 'I' (capital i) doesn't get detected in my plates, please check the given example.

Any thoughts on this?

Thanks!

Kind regards,
Ben

Is it possible to offer an alternative license to AGPL? Any license which doesn't require open sourcing proprietary code or offering an optional commercial license would fit our requirements (free users use AGPL, commercial users pay $ and keep their code closed source).

Thanks, Daniel.

Hello!
I am trying to train openalpr for Bulgarian plates.

I have a cropped images with license plates and I am running classifychars. It gives me this error:

OpenCV Error: Bad flag (parameter or structure field) (Unrecognized or unsupported array type) in cvGetMat, file /opencv/opencv-2.4.8/modules/core/src/array.cpp, line 2482
terminate called after throwing an instance of 'cv::Exception'
what(): /opencv/opencv-2.4.8/modules/core/src/array.cpp:2482: error: (-206) Unrecognized or unsupported array type in function cvGetMat

Please advice.

Hello,

To compile on ubuntu 12.04, I had to add #include "resultiterator.h" in ocr.h:

/home/philippe/work/openalpr/src/openalpr/ocr.cpp: In member function ‘void OCR::performOCR(std::vector<cv::Mat>, std::vector<cv::Rect_<int> >)’:
/home/philippe/work/openalpr/src/openalpr/ocr.cpp:84:32: error: invalid use of incomplete type ‘struct tesseract::ResultIterator’
/usr/include/tesseract/baseapi.h:83:7: error: forward declaration of ‘struct tesseract::ResultIterator’

And then the binary segfaulted in ocr.cpp around line 100;

const char* choice = ci.GetUTF8Text();
postProcessor->addLetter(*choice, j, ci.Confidence());

It segfault because choice is value 0x0. If I comment the whole "post-process" block then it still work and gives decent results.

I did install opencv manually and for tesseract I did "apt-get install libtesseract-dev", which is version 3.02.01-2.

Any idea about why I get the segfault?

Hi,

I am a newbie and would love to use openalpr to read some license plates into a database. Unfortunately, I am having trouble with the region detection so I am wondering if I am doing something wrong or whether it is just not good.

I am running the command "alpr -d license.jpg -r <runtime_dir>" to try to get the region into the output. The license plate is reading fine but I get (no region)

plate0: 6 results -- Processing Time = 245.592ms.
- 478HM2 confidence: 89.4472 template_match: 0
- 478HN2 confidence: 81.5051 template_match: 0
- 478H2 confidence: 80.1997 template_match: 0

if I add the json flag, I see:

[{"plate":"478HM2","confidence":89.447197,"matches_template":0,"region":"","region_confidence":0,"processing_time_ms":100.312790,"coordinates":[{"x":405,"y":311},{"x":545,"y":308},{"x":545,"y":385},{"x":405,"y":385}],"candidates":............

So it looks for this plate, it just can't get the region. (It's Massachusetts but the plate is on the car and somewhat obscured.)

I also tried on a VT and NH plate and it guessed "GA" and "OK".

I attached the last 2 here in case someone is kind enough to advise and interested in seeing them. Thanks in advance!

The angle of the edge of the plate can be better determined by examining the angle of the characters. In the picture attached, notice that the first pass (charanalysis) identified a few valid characters. Looking at the edges of the U character, we know the angle of the edge that we're looking for. It should be slanted backwards. Without this knowledge, the edge detector picked a vertical line instead.

When printing as json the resulting output is invalid:

{"epoch_time":1408627909,"processing_time_ms":67.107000,"results":[]}
No license plates found.

The message "No licence plate found" should only be printed in the console, not to json.

After loading and extracting openalpr.windows.zip, I run

alpr.exe sample.jpg

I get the following error

--(!)Runtime directory 'C:/projects/openalpr/src/../runtime_data/' does not exist!
--(!)Error loading classifier
Tesseract couldn't load any languages!
OpenCV Error: Assertion failed (scaleFactor > 1 && image.depth() == CV_8U) in cv::CascadeClassifier::detectMultiScale, file src\cascadedetect.cpp, line 1081

Sure, this workaround helps

xcopy /e runtime_data\ c:\projects\openalpr\runtime_data

but a relative path would be nicer I think. Especially for users without writing permission to c:.

Running OpenALPR in multiple forked processes does increase the performance significantly. However, the multi-threading within a single process never increases performance.

Currently, OpenALPR has a multi-threading function implemented (controllable by setting the number of cores in openalpr.conf). But it does not work. Threads are instantiated correctly, but they are just as slow if not slower than if the processing was run serially.

I am testing openalpr on iOS, and noticed via profiling that a chunk of time (~ 3-5 sec on an iPhone 4S) is spent in instantiating the cv::BRISK object for the FeatureMatcher extractor. How do you feel about implementing a caching mechanism (serialization?), since it appears that the values for instantiation are static? Admittedly, I have never done serialization in C++, but I could look into it if you would like.

Hello,

The code uses new/delete pretty much everywhere, I think:

1. Most of them are not needed at all, e.g both postProcessor and tesseract at https://github.com/openalpr/openalpr/blob/develop/src/openalpr/ocr.cpp#L26-28 could simply be data members instead, no new/delete needed.
2. Those who really need to be a pointer can use std::scoped_ptr or std::shared_ptr depending on the need.

Another example here: Silex@b548734

All output contains "template_match: 0" even when the -t option is not given. This should be removed in those cases.

Some method of taking note of the positions of plates that are recognised and forming a mask to only analyse those areas would be a great feature.

Cheers

Steve

Setup a system that automatically modifies parameters, then runs a benchmark. The result should be piped into MOE to get the statistically optimal values.

For example, plate corner detection uses a number of tuneable weights. A benchmark could vary these values and re-run the benchmark and have MOE evaluate the best settings.

Many other ALPR machine-learning parameters could be analyzed in a similar way.

https://github.com/Yelp/MOE

The shadow bisecting the plate image causes errors in the OCR output. In this case, the "J" is believed to be a "3"

Hi!

I'm trying to build this on Raspbian, but 'make' fails. I'm sort of new to github and how to properly format a bugreport/problem, so bare with me :)

Here is the output from "make" after following the instructions for installation:
http://pastebin.com/dK90kie5

Any suggestions?