zxm97/RetinaFace_LPRNet_TensorRT

TensorRT inference (Python and C++) for Chinese single-line and double-line license plate detection and recognition. Optimized for Jetson Nano. 车牌检测，车牌识别，支持单层和双层车牌识别，有 TensorRT Python 和 C++ 的demo，适合 Jetson Nano 运行。

Python

English, 中文

Plate type supported

Chinese license plate (single-line and double-line)

Environment

PC

Windows 10

python 3.7.10

CUDA 11.7

cuDNN 8.6.0.163

TensorRT 8.5.1.7

OpenCV 4.6.0 (build from source)

CMake 3.26.5

Visual Studio 16 2019

Jetson Nano

python 3.6.9

Jetpack 4.6.2

CUDA 10.2.89

cuDNN 8.2.1.32

TensorRT 8.2.1.8

OpenCV 4.5.2 with GStreamer support

Requirements

PC

albumentations==1.3.1

imutils==0.5.4

matplotlib==3.3.4

numpy==1.21.6

opencv_python==4.8.1.78

opencv_python_headless==4.8.1.78

Pillow==9.5.0

pycuda==2021.1+cuda115

pytorch_lightning==1.8.0

PyYAML==5.4.1

task==0.2.5

termcolor==2.3.0

torch==1.13.1+cu116

torchvision==0.14.1+cu116

Jetson Nano

albumentations==1.3.1

imutils==0.5.4

matplotlib==3.3.2

numpy==1.16.1

Pillow==10.1.0

pycuda==2021.1

pytorch_lightning==1.5.0

PyYAML==6.0.1

task==0.2.5

termcolor==1.1.0

torch==1.6.0a0+b31f58d

torchvision==0.9.0a0+01dfa8e

How to use

Run PyTorch demo on PC:

run demo_torch.py

Run TensorRT demo on PC (python):

step 1

clone the project and follow the instructions to build Retinaface trt engine:

https://github.com/zxm97/Pytorch_Retina_License_Plate_trt

step 2

clone the project and follow the instructions to build LPRNet trt engine:

https://github.com/zxm97/license-plate-recoginition_trt

step 3

run demo_trt.py / demo_trt_fpn_reduced.py / demo_trt_fpn_reduced_async.py

demo_trt.py is for original RetinaFace model
demo_trt_fpn_reduced.py is for the modified RetinaFace model (1 layer of FPN is reduced)
demo_trt_fpn_reduced_async.py is for the modified RetinaFace model (1 layer of FPN is reduced). Speed up by dividing the task into threads.

Run TensorRT demo on PC (C++):

step 1

clone the project and run gen_wts_for_tensorrtx.py to get weight map file: https://github.com/zxm97/Pytorch_Retina_License_Plate_trt

step 2

open xxx\Project_cpp\CMakeLists.txt，modify OpenCV_DIR and TRT_DIR。

open CMake (GUI)

set source code directory : xxx/Project_cpp

set binaries directory : xxx/Project_cpp/build

click Configure, Generate and Open Project

step 3

select Release x64, build solution for decodeplugin, create xxx\Project_cpp\build\Release\decodeplugin.lib

step 4

build solution for retina_mnet_plate

move the weight map file in step 1 to xxx\Project_cpp\build\Release

open Windows PowerShell, run the following command to build RetinaFace trt engine:

./retina_mnet_plate.exe -s

step 5

clone the project and follow the instructions to build LPRNet trt engine:

https://github.com/zxm97/license-plate-recoginition_trt

move LPRNet trt engine file to xxx\Project_cpp\build\Release

step 6

build solution for demo

run the following command to detect and recognize plates on images:

./demo.exe -d

Run TensorRT demo on Jetson Nano (python):

step 1

clone the project and follow the instructions to build Retinaface trt engine (use the modified model with 1 FPN layer reduced):

https://github.com/zxm97/Pytorch_Retina_License_Plate_trt

step 2

clone the project and follow the instructions to build LPRNet trt engine:

https://github.com/zxm97/license-plate-recoginition_trt

step 3

run demo_trt_jetson_fpn_reduced_async.py

Speed up by dividing the task into threads
To lower CPU usage, use hardware acceleration for decoding.
Up to 20 frames per second for a 1258 x 684 h.264 video

References

https://github.com/gm19900510/Pytorch_Retina_License_Plate

https://github.com/ChHanXiao/license-plate-recoginition

https://github.com/1996scarlet/faster-mobile-retinaface

https://github.com/wang-xinyu/tensorrtx

https://github.com/SunlifeV/CBLPRD-330k

https://github.com/yxgong0/CRPD

https://gist.github.com/jkjung-avt/86b60a7723b97da19f7bfa3cb7d2690e