Number Plate Recognition

A yolov4 tiny model trained using Darknet and converted to pytorch to improve efficency

Environment

Developed in an Arch environment but should work on Windows and Mac as well

Screenshots

Installation

Pytorch

Install Dependencies

Common

conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi typing_extensions future six requests dataclasses

On Linux

# Add LAPACK support for the GPU if needed
conda install -c pytorch magma-cuda102  # or [ magma-cuda101 | magma-cuda100 | magma-cuda92 ] depending on your cuda version

On MacOS

# Add these packages if torch.distributed is needed
conda install pkg-config libuv

On Windows

# Add these packages if torch.distributed is needed.
# Distributed package support on Windows is a prototype feature and is subject to changes.
conda install -c conda-forge libuv=1.39

Get the PyTorch Source

git clone --recursive https://github.com/pytorch/pytorch
cd pytorch
# if you are updating an existing checkout
git submodule sync
git submodule update --init --recursive

Install PyTorch

On Linux

export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
python setup.py install

On macOS

export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install

More installation instructions

OpenCV

On Linux

install deoendencies

sudo apt-get update
sudo apt-get install build-essential
sudo apt-get install cmake git libgtk2.0-dev pkg-config libavcodec-dev libavformat-dev libswscale-dev

optional

sudo apt-get install python-dev python-numpy libtbb2 libtbb-dev libjpeg-dev libpng-devlibtiff-dev libjasper-dev libdc1394-22-dev

Build

mkdir ~/src
cd ~/src
git clone https://github.com/opencv/opencv.git
cd opencv
mkdir build && cd build

cmake -D CMAKE_BUILD_TYPE=RELEASE \
      -D CMAKE_INSTALL_PREFIX=/usr/local \
      -D INSTALL_PYTHON_EXAMPLES=ON \
      -D INSTALL_C_EXAMPLES=ON ..
make -j$(nproc)
sudo make install

Openalpr

Install openALPR for Ubuntu (Much easier and had a higher success rate)

# Install prerequisites
sudo apt-get install libopencv-dev libtesseract-dev git cmake build-essential libleptonica-dev
sudo apt-get install liblog4cplus-dev libcurl3-dev

# If using the daemon, install beanstalkd
sudo apt-get install beanstalkd

# Clone the latest code from GitHub
git clone https://github.com/openalpr/openalpr.git

# Setup the build directory
cd openalpr/src
mkdir build
cd build

# setup the compile environment
RUN cmake -D WITH_BINDING_JAVA=OFF \
  -D WITH_DAEMON=OFF \
  -D WITH_GPU_DETECTOR=ON \
  -D CMAKE_INSTALL_PREFIX:PATH=/usr \
  -D CMAKE_INSTALL_SYSCONFDIR:PATH=/etc ..

# compile the library
make -j4

# Install the binaries/libraries to your local system (prefix is /usr)
sudo make install

# Test the library
wget http://plates.openalpr.com/h786poj.jpg -O lp.jpg
alpr lp.jpg

Once you've completed the steps, go to the git repo you've cloned to your machine during the above installation and install the python bindings. runtime_data folder is also available in this repo, which will be helpful for next step.

cd openalpr/src/bindings/python/
sudo python3 setup.py install

Usage

python main.py -cfgfile <path to cfg file>\
               -weightfile <path to weight file>\
               -namesfile <path to classes.names file>\
               -output <name for the output video>\
               -source <path to video if you want to use video else index for webcams>

Up vote

If you found this useful please upvote this repository

Opimizations

Optimize openalpr for some perfomance gains

Modify file /etc/openalpr/openalpr.conf

; This configuration file overrides the default values specified 
; in /usr/share/openalpr/config/openalpr.defaults.conf
hardware_acceleration = 1
gpu_id = 0
gpu_batch_size = 10

Modify file /usr/share/openalpr/openalpr.defaults.conf

max_detection_input_width = 800
max_detection_input_height = 600

detector = lbpgpu

skip_detection = 1

max_plate_angle_degrees = 30

change the resolution in cfg/yolov4.cfg

[net]
# Testing
#batch=1
#subdivisions=1
# Training
batch=64
subdivisions=16
width=608      #change the height and width lower resolutions
height=608     #improve perfomance and higher resolutions 
channels=3     #improve accuracy
momentum=0.9
decay=0.0005
angle=0
saturation = 1.5
exposure = 1.5
hue=.1

Perfomance

• GPU Tesla T4 (without OCR)

Res	FPS (avg)
800x800	56
640x640	68
608x608	72
480x480	91
416x416	98
320x320	106

I couldn't get Openalpr to build on colab. With ocr FPS will tank ~10fps

Added numba to optimize the inference loop a little bit its not much but its something there was a bit of a perfomance boost for me while running in an Intel Pentium N3710 especially in nms from 278avg to 177avg.

If you are much more familiar with numba you can help by optimizing it much more if possible i dont own a GPU currently so cant test anything out.

Todo

1. Save cropped plate images

Credits

Darknet YOLOv4 - Neural Networks for Object Detection (Windows and Linux version of Darknet )

pytorch-YOLOv4 PyTorch ,ONNX and TensorRT implementation of YOLOv4

Pytorch Tensors and Dynamic neural networks in Python with strong GPU acceleration

OpenCV Open Source Computer Vision Library

Openalpr Automatic License Plate Recognition library