ALPR-with-Yolo-v4

ALPR with YOLOv4 is an advanced Automatic License Plate Recognition (ALPR) system that leverages the powerful YOLOv4 (You Only Look Once) one-stage object detection framework. It can efficiently and accurately detect and recognize vehicle license plates in real-time.

About Darknet : http://pjreddie.com/darknet/

Download Model

• .weights : https://drive.google.com/file/d/1b3rYgP48z_NGvSuNoMKDvxXzYmray_Qr/view?usp=sharing
• .mlmodel : https://drive.google.com/file/d/1eREdAVoiOVlAiPOxv5c_Sc5EBPybOIUq/view?usp=sharing

Classes

• car
• license_plate

Training

Labeling Tool : https://github.com/AlexeyAB/Yolo_mark

Darknet (Yolov4) : https://github.com/AlexeyAB/darknet

Cloud Service	GPU	Traing Data	Training Iterations	Time
GCP(Google Cloud Platform)	Nvidia Tesla P100	Over 2600 images	4000 iterations	5h

./darknet detector train data/obj.data cfg/yolov4_ANPR.cfg yolov4.conv.137 -gpu 0

Usage (test)

1. git clone https://github.com/AlexeyAB/darknet
2. cd darknet
3. Configure Makefile according to your environment: vi Makefile

GPU=0        # Change to 1 if using GPU
CUDNN=0      # Change to 1 if using cuDNN (NVIDIA)
CUDNN_HALF=0
OPENCV=0     # Change to 1 if using OpenCV
AVX=0
OPENMP=0
LIBSO=1      # Generate libdarknet.so

...
...

4. make

• Required packages: make, gcc, pkg-config (if not installed, use sudo apt-get install … to install)

5. Download data/*, cfg/yolov4-ANPR.cfg, and backup/yolov4-ANPR.weights

image

./darknet detector test data/obj.data cfg/yolov4-ANPR.cfg backup/yolov4-ANPR.weights data/(이미지파일.jpg)

Make sure to use .jpg images

video

./darknet detector demo data/obj.data cfg/yolov4-ANPR.cfg backup/yolov4-ANPR.weights data/(동영상파일.mp4)

webcam

./darknet detector demo data/obj.data cfg/yolov4-ANPR.cfg backup/yolov4-ANPR.weights

Example

Prediction Image

./darknet detector test data/obj.data cfg/yolov4-ANPR.cfg backup/yolov4-ANPR.weights data/testfile.jpg

Loading weights from backup/yolov4-ANPR.weights...
 seen 64, trained: 256 K-images (4 Kilo-batches_64)
Done! Loaded 162 layers from weights-file
data/testfile.jpg: Predicted in 9325.005000 milli-seconds.
car: 63%
car: 98%
license_plate: 96%
car: 47%
car: 61%
car: 30%

Prediction Video

• ./darknet detector demo data/obj.data cfg/yolov4-ANPR.cfg backup/yolov4-ANPR.weights data/testvideo.jpg
• python darknet_video.py

Demo Video Link (1) : https://drive.google.com/file/d/1DGmF2bwtDMe1y-wNuv_YT827Vr6Y8Q2m/view?usp=sharing

Demo Video Link (2) : https://drive.google.com/file/d/1nJjIQFcrYRYSJ0n9FK0-x_Fk6HrULsZY/view?usp=sharing

References

• Papers
  • You Only Look Once : Unified, Real-Time Object Detection
  • YOLO9000 : Better, Faster, Stronger
  • YOLOv3 : An Incremental Improvement
  • YOLOv4：Optimal Speed and Accuracy of Object Detection
• Keras-yolov3
  • weights to h5 : https://github.com/qqwweee/keras-yolo3/blob/master/convert.py
  • weights to mlmodel : https://gist.github.com/TakaoNarikawa/aef13571eec97d78603688eef05b5389
  • Mish : https://qiita.com/TakaoNarikawa/items/e4521fd8c7a522e9d4fd
• Core ML
  • https://gist.github.com/TakaoNarikawa
  • https://github.com/Ma-Dan/YOLOv3-CoreML

Presentation

• 발표자료 : https://drive.google.com/file/d/1yhhIZ0ZU5MIZZar-WTBGgEHwLkOHcewe/view?usp=sharing

• 발표영상 : https://www.youtube.com/watch?v=H3-SVf0Ps4c

• Capstone Design : https://github.com/kwanghoon/CapstoneDesign