ALPR system on xavier

[ozlematiz1637]

Hello to everyone. I am making a license plate recognition system through Xavier developer. I am processing on the real image. I have a problem. If the plate is in the image, it tries to be read from the first frame. But it misreads the license plate in the first frame while in motion. When the vehicle stops, it gives 95% accurate results.
My question is: How can I choose the correct one from the plate data coming from the frames? I thought I’d accept the most repeated one as correct. And I added the plates to a list. I want the counter to accept the most repeated one as correct. But it seems like a bit of a complicated process to me.

If the vehicle is in motion, I want it to not do anything. When the vehicle stops, the system should read the plate. How can I do that? I would be very happy if anyone can help.

Thanks

Developer kit: Nvidia Jetson Xavier AGX
OS:Ubuntu 20.04
Object Detection Models: Yolov5

[AastaLLL]

Hi,

Suppose you have a detector for the license plate as well as a classifier for the character.

A possible way is to apply object tracking to the detector first.
So you will know the corresponding object cross different frames.

With this information, you can count the recognition label generated from the same object to reach your goal.

Thanks.

[ozlematiz1637]

@AastaLLL thank you for your reply. I got this. But I’m confused on how to do it. Can you elaborate a bit? Thank you

[AastaLLL]

Hi,

Could you share more information about your source?

Suppose you use YOLOv5 for license plate detection.
Is there another classifier to recognize the number?

More, which library do you use?
Thanks.

[ozlematiz1637]

I am using the yolov5 model. mongodb as database.
I accept the most repetitive plate as correct

[ozlematiz1637]

torch
PIL
cv2
numpy
pandas
time
datetime
pymongo
os

[AastaLLL]

Hi,

Maybe you can try to calculate the motion between two adjacency frames.
The motion should be very different when a car is moving or stopping.

Thanks.