SOD-based School Zone Vehicle License Plate Detection

Project Overview

Developed Small Object Detection (SOD) technology to enhance license plate recognition in school zones, addressing the limitations of YOLO-based object detection models.

** about SOD (Our company and University developed this technology )

bookmark

Key Responsibilities

• Collected and preprocessed vehicle license plate data.
• Built a Korean vehicle license plate generator using OpenCV and Python.
• Set up a Kubeflow-based environment for license plate recognition model development.
• Managed AI project configurations using Cookiecutter, MinIO, DVC, and Git.

Achievements

• Although the project wasn't adopted initially, it contributed to winning another research project in the following year.
• Increased model accuracy to 98% by generating and augmenting license plate datasets.

Key Learnings and Insights

This project provided valuable experience in computer vision, small object detection (SOD), and real-world AI implementation, particularly for license plate recognition in school zones.

1. Advancing AI for Traffic Monitoring and Enforcement
   • Initially researched illegal parking and lane violation detection (AI Police) before expanding into license plate recognition using SOD technology.
   • Discovered how small object detection techniques can significantly enhance object recognition accuracy in real-world scenarios.
2. Deepening Expertise in Computer Vision and Dataset Engineering
   • Explored various methods to improve vision-based license plate recognition accuracy, specifically for low-resolution and distant objects.
   • Developed a Korean vehicle license plate generator using OpenCV and Python, enabling robust dataset augmentation.
   • Learned the critical role of high-quality datasets in improving model performance, realizing the gap between research datasets and real-world data.
3. Impact and Long-Term Contributions
   • Although the initial project wasn’t adopted, the research and development efforts led to securing a follow-up research project, demonstrating the long-term value of foundational AI research.
   • Achieved a 98% model accuracy improvement through dataset generation, augmentation, and small object detection techniques.

This project reinforced the importance of dataset quality, real-world application challenges, and advanced small object detection methods, while also deepening my expertise in AI-driven traffic monitoring solutions.