Building and Testing 'openalpr' on Jetson TX2

Mar 9, 2018

I read about openalpr a while ago. Recently I was requested to integrate license plate recognition function into our TX2 product. ‘openalpr’ came up as my go-to solution for the task.

There seems to be a pre-built ‘openalpr’ apt package available, but it has dependency on opencv apt packages. I didn’t want ‘apt install’ to mess up with the opencv-3.4.0 I built and installed on JTX2, so I chose to build ‘openalpr’ from source. Here’s how I’ve done it.

Prerequisite

Build and install opencv-3.4.0 on the target JTX2. Reference: How to Install OpenCV (3.4.0) on Jetson TX2

Step-by-step

In addition to ‘opencv’, ‘leptonica’ and ‘tesseract’ are also required for ‘openalpr’. The build process is broken down into several steps below. Reference: Compile openalpr The Harder Way (Compile all dependencies manually)

Step #1: build and install ‘leptonica’

I installed the latest version of ‘leptonica’: 1.75.3.

First, install dependencies.

$ sudo apt-get install autoconf automake libtool
$ sudo apt-get install autoconf-archive pkg-config
$ sudo apt-get install libpng-dev libjpeg8-dev libtiff5-dev zlib1g-dev
$ sudo apt-get install libicu-dev libpango1.0-dev libcairo2-dev

I’d also set TX2 to maximum performance mode in order to speed up the building process.

$ sudo nvpmodel -m 0
$ sudo ~/jetson_clocks.sh

Then download and install ‘leptonica’.

$ mkdir -p ~/src
$ cd ~/src
$ wget http://www.leptonica.com/source/leptonica-1.75.3.tar.gz
$ tar xzvf leptonica-1.75.3.tar.gz
$ cd leptonica-1.75.3/
$ ./configure --prefix=/usr/local
$ make -j4
$ sudo make install

Step #2: build and install ‘tesseract’

I installed the latest stable version of ‘tesseract’: 3.05.01. Reference: Compiling (tesseract)

$ cd ~/src
$ git clone https://github.com/tesseract-ocr/tesseract.git
$ cd tesseract/
$ git checkout 3.05
$ ./autogen.sh 
$ ./configure --prefix=/usr/local
$ make -j4
$ sudo make install

In addition to the libraries and excutbales, I also had to download the language data files needed for OCR. More specifically, I downloaded and installed the ‘English’ and ‘Chinese - Traditional’ data files. Reference: Data Files (tesseract)

$ cd ~/Downloads/
$ wget https://github.com/tesseract-ocr/tessdata/raw/3.04.00/eng.traineddata
$ wget https://github.com/tesseract-ocr/tessdata/raw/3.04.00/chi_tra.traineddata
$ sudo cp eng.traineddata /usr/local/share/tessdata/
$ sudo cp chi_tra.traineddata /usr/local/share/tessdata/

Following tesseract’s documentation I put the following into my ~/.bashrc. Log out and log in again to let the TESSDATA_PREFIX environment variable take effect.

export TESSDATA_PREFIX=/usr/local/share

Step #3: build and install ‘openalpr’

Install dependencies and clone the ‘openalpr’ repo.

$ sudo apt-get install curl libcurl4-openssl-dev
$ sudo apt-get install liblog4cplus-dev
$ cd ~/src
$ git clone https://github.com/openalpr/openalpr.git
$ cd openalpr

‘openalpr’ needs to be compiled with gcc ‘-std=c++11’ flag. So I modified src/CMakeLists.txt as follows. I put a copy of my modified src/CMakeLists.txt here, for reference.

--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -5,6 +5,8 @@ set(CMAKE_BUILD_TYPE RelWithDebugInfo)
  
 cmake_minimum_required (VERSION 2.6)

+set (CMAKE_CXX_STANDARD 11)
+
 # Set the OpenALPR version in cmake, and also add it as a DEFINE for the code to access
 SET(OPENALPR_MAJOR_VERSION "2")
 SET(OPENALPR_MINOR_VERSION "3")

I also needed to make some modification to src/daemon.cpp, otherwise compilation of that source file would fail. Again I put a copy of my modified src/daemon.cpp here, for reference. Check out this StackOverflow thread for more information about this modification.

--- a/src/daemon.cpp
+++ b/src/daemon.cpp
@@ -156,7 +156,8 @@ int main( int argc, const char** argv )
     log4cplus::SharedAppenderPtr myAppender(new log4cplus::RollingFileAppender(logFile));
     myAppender->setName("alprd_appender");
     // Redirect std out to log file
-    logger = log4cplus::Logger::getInstance(LOG4CPLUS_TEXT("alprd"));
+    auto tmp_logger = log4cplus::Logger::getInstance(LOG4CPLUS_TEXT("alprd"));
+    logger = tmp_logger;
     logger.addAppender(myAppender);
  
  
@@ -167,7 +168,8 @@ int main( int argc, const char** argv )
     //log4cplus::SharedAppenderPtr myAppender(new log4cplus::ConsoleAppender());
     //myAppender->setName("alprd_appender");
     // Redirect std out to log file
-    logger = log4cplus::Logger::getInstance(LOG4CPLUS_TEXT("alprd"));
+    auto tmp_logger = log4cplus::Logger::getInstance(LOG4CPLUS_TEXT("alprd"));
+    logger = tmp_logger;
     //logger.addAppender(myAppender);
  
     LOG4CPLUS_INFO(logger, "Running OpenALPR daemon in the foreground.");

Here are the commands for building ‘openalpr’.

$ cd ~/src/openalpr/src
$ mkdir build
$ cd build
$ cmake -D WITH_GPU_DETECTOR=ON ..
$ make -j4
$ sudo make install

Step #4: testing

In order to achieve better performance of ‘openalpr’ by utilizing the GPU on TX2, I had to modify the config file. More specifically, I made a copy of openalpr’s default config file and changed detector from lbpcpu to lbpgpu, along with other minor modifications. I put a copy of my modified config file here, for reference.

$ mkdir -p ~/src/openalpr/test
$ cd ~/src/openalpr/test
$ cp /usr/local/share/openalpr/config/openalpr.defaults.conf tx2.conf
$ vim tx2.conf

Finally I used a picture from my mobile phone for testing.

A photo of ASM-9197

Here’s the result…

nvidia@tegra-ubuntu:~/src/openalpr/test$ alpr --country us --config tx2.conf ASM-9197.jpg
--(!)Loaded CUDA classifier
plate0: 10 results
    - ASM9197    confidence: 88.5777
    - ASM9I97    confidence: 85.3164
    - A5M9197    confidence: 81.3223
    - A5M9I97    confidence: 78.0611
    - ASH9197    confidence: 76.4136
    - ABM9197    confidence: 76.0675
    - AM9197     confidence: 75.7315
    - ASH9I97    confidence: 73.1524
    - ABM9I97    confidence: 72.8063
    - AM9I97     confidence: 72.4703