Automatic License Plate Recognition Implementation

[mohammedalsaeed-uob]

• Decide on implementation (https://github.com/computervisioneng/real-time-number-plate-recognition-anpr) or (https://github.com/licesonw/anpr-sagemaker-custom?tab=readme-ov-file#automated-number-plate-recognition-with-amazon-sagemaker)
• Implement Chosen Solution

[mohammedalsaeed-uob]

Added new resource to use in implementation. https://github.com/Muhammad-Zeerak-Khan/Automatic-License-Plate-Recognition-using-YOLOv8

[mohammedalsaeed-uob]

Brainstormed a few tasks:

1. Add function to delete snapshots captured by Unregistered Cars Consumer
2. Add function to retain snapshots captured by Yellow Lane Violation Consumer
3. Create Folders for Separate Violations in S3 Violation Bucket
4. Create DynamoDB table for Yellow Lane Violation License Plate Storage
5. Edit function to check for unregistered car violation

(will add these as their own issues)

[mohammedalsaeed-uob]

We successfully developed Consumer 1.

• It takes the stream as input from Kinesis Video Streams.
• A car and license plate detector model on the captures images of license plates and saves them to S3
• A lambda function takes the image from S3 and runs it against Amazon Textract to get the license plate number
• The same lambda function stores the extracted license plate number into a Dynamo DB table.

[mohammedalsaeed-uob]

Inquiry @maramesk @cerashdan

The model is alright, but it's not the most accurate: It sometimes captures extra characters. We have the option of training it on Bahraini cars. Our question is: Is that a priority? There are two key points to consider here

• This model is reused throughout all our consumers to detect cars and their license plates.
• We're not sure if it's what we should focus our efforts on right now.

Should we focus some of our efforts on training the model now? Or should we come back to it at a later time when we've made more the other consumers (for the other violations)?

P.s. After you respond we will close this issue, since the task itself is complete. Just leaving it opened to discuss this point.

[maramesk]

@mohammedalsaeed-uob The returning students from the Amakin Challenge (currently working in BQA) have implemented accurate plate detection on Bahraini plate numbers, reach out to them to check their approach. They did not have to train their own model on Bahraini plates and I don't think its feasible for you to do so either.

[mohammedalsaeed-uob]

@maramesk Will do, thank you.

[cerashdan]

Adding few more pointers to guide you:

1. When solving computer vision problems, think about your scene and try to eliminate unnecessary information. If you have a scene with many cars, try the following:

• Detect cars in the scene
• Save the car bounding box in an array (n m k numpy array where n is the number of cars and m * k is the resized cropped car)
• Filter out the interest region, Bahraini plates have a red flag, dark edge and a white background. Find methods that can detect this area.
• Once you find this area, crop only the license plate. then run your ML model to detect the digits.

Using textract for every frame is unpractical as you are processing a video. This problem could easily be solved using ML approaches running on the consumer itself.

[cerashdan]

Automatic License/Number Plate Recognition (ANPR/ALPR) is a solved problem in Computer Vision, have a look at:

https://pyimagesearch.com/2020/09/21/opencv-automatic-license-number-plate-recognition-anpr-with-python/

More specifically this section: Lines 30 and 31 perform a blackhat morphological operation to reveal dark characters (letters, digits, and symbols) against light backgrounds (the license plate itself). As you can see, our kernel has a rectangular shape of 13 pixels wide x 5 pixels tall, which corresponds to the shape of a typical international license plate

This blog post above, shows you how to solve this problem purely using Computer Vision approaches without using ML Models/Training.

[mohammedalsaeed-uob]

@cerashdan Thank you for sharing these invaluable pointers and resources. I would like to point out that we are not using Textract on every frame, so no need to worry about that. After adjusting the code, it generally works much better now. Thank you once again.

[mohammedalsaeed-uob]

Currently migrating our functions and code to SST/CDK per the following: -edit lambda to get queue URL on stack creation (husain) -edit lambda to get permissions with iam role (husain) ✅ -create ec2 /computer stack (mohammed) -create ssm/run command and userdata (mohammed) -edit S3 notification to trigger correct lambda (husain)✅

then we can close this issue.

[HusainAdel]

87 and #88 branches for dyanamo, s3, lambda functions