This scientific research project, conducted at Duy Tan University (DTU), aims to develop a robust system for detecting traffic violations using YOLOv5 models in conjunction with OpenCV.

Link: https://youtu.be/ddhAg65acd0

The project leverages three distinct YOLOv5 models to address different components of traffic violation detection:

• Vehicle Detection: Identifies vehicles on the road.
• Helmet Detection: Identifies helmet on the road.
• Plate Detection: Extracts the license plate from the detected vehicle.
• Number Plate Detection: Extracts the numeric characters from the license plate.

The detection pipeline involves the following steps:

1. Vehicle Detection: Locating the vehicle on the road.
2. Helmet Detection: (If applicable) Identifying whether the rider is wearing a helmet.
3. License Plate Detection: Isolating the vehicle’s license plate.
4. Number Extraction: Extracting numeric values from the license plate.

Given that real-world images often lack the clarity of training images, pre-processing is crucial. We apply a blurring technique to improve the performance of the Number Plate Detection model.

• Before Pre-processing:

  

• After Pre-processing:

  

Model training is performed on Kaggle using YOLOv5. More details can be found in the Kaggle Notebook.

For a deeper dive into YOLOv5, refer to the official documentation.

A pipeline integrates the three models, where each model's output serves as the input for the next.

1. Vehicle Detection:

   • Input: Images, Mask, Separation Line (for Lane Encroachment Detection).
   • Output: Images highlighting the detected violations.

   

2. Plate Detection:

   • Output: Cropped images containing the vehicle's license plate.

   

3. Number Plate Detection:

   • Output: Extracted numeric values, e.g., 59N12345, 76E152202, ...

You can view and explore the dataset used in this project on Kaggle.

Pre-trained weights for the helmet detection model can be downloaded here.