PlateTrack AI

An end-to-end, production-ready AI system for automatic vehicle license plate detection and text extraction, powered by YOLO, OpenCV, and EasyOCR.

PlateTrack AI takes you from model training to real-world deployment: it detects plates in images and videos, performs OCR with high accuracy, and delivers results through a deployed interactive web application.

Deployed Project

• Status: Successfully deployed
• Live Demo: https://platetrack-ai-hhdrpbs8arknapjcbsdt54.streamlit.app/

Project Overview

PlateTrack AI is a completed end-to-end Automatic Number Plate Recognition (ANPR) project with three major parts:

1. Model development workflow in notebook for dataset preparation, training, and evaluation.
2. Streamlit UI for inference on uploaded media (images/videos).
3. Deployment of the application for live web access.

Key Features

• License plate detection using trained YOLO model (models/best.pt)
• Image inference with bounding box visualization
• Video inference frame-by-frame with bounding box overlays
• OCR extraction of detected plate text from image regions (EasyOCR)
• Browser-friendly video output conversion using FFmpeg
• Streamlit frontend with project info, stack, and use cases
• Deployed web application for real-world usage and demo access

Tech Stack

• Python
• Streamlit
• Ultralytics YOLO
• OpenCV
• EasyOCR
• FFmpeg (via imageio-ffmpeg)
• NumPy, Pandas, Matplotlib

Repository Structure

Automatic-Car-Numberplate-Recognition-System/
|-- .gitignore
|-- yolo_application.py
|-- requirements.txt
|-- README.md
|-- LICENSE
|-- datasets/
|   |-- annotations/
|   |-- images/
|   `-- demo.mp4
|-- models/
|   `-- best.pt
|-- notebooks/
|   `-- Car Numberplate Detection.ipynb
|-- output/
|   |-- demo.mp4
|   |-- demo_web.mp4
|   `-- images.jpeg
|-- temp/
|   |-- demo.mp4
|   `-- images.jpeg
|-- test/
|   |-- demo.mp4
|   `-- images.jpeg
`-- yoloenv/

Dataset

• Total images: 433 (datasets/images)
• Total annotations: 433 Pascal VOC XML files (datasets/annotations)
• Annotated images: 433 (0 without annotations)
• Total bounding boxes: 471
• Classes: licence (471 instances)
• Boxes per image: min 1, max 6, mean 1.09, median 1
• Top image resolutions (sample): 400×300 (64), 400×267 (50), 400×225 (47)
• Parsing errors while analyzing: 0

This dataset was used to train the YOLO model at models/best.pt.

How It Works

1) Detection

• YOLO model predicts plate bounding boxes (x1, y1, x2, y2) and confidence.
• Boxes and confidence text are drawn on image/video frames.

2) OCR

• For each detected plate in images, ROI is cropped from the bounding box.
• EasyOCR reads text from ROI.
• Extracted text is shown in Streamlit.

3) Video Preview Conversion

• Processed video is converted to web-friendly MP4 (H.264 + yuv420p + faststart).
• This improves compatibility in browser playback.

Model Development (Notebook)

The notebook notebooks/Car Numberplate Detection.ipynb includes:

• XML annotation parsing and dataframe creation
• Pascal VOC to YOLO coordinate conversion
• Train/validation/test split preparation
• YOLO training and metric plotting
• Image-level detection testing
• OCR integration experiments
• Video detection workflow

Installation

From project root:

pip install -r requirements.txt

Run the Application

streamlit run yolo_application.py

Then open the Streamlit URL shown in terminal and upload:

• Image: .jpg, .jpeg, .png, .bmp
• Video: .mp4, .avi, .mov, .mkv

Example Output

• Input image/video with detected license plate boxes
• Confidence labels on each detection
• OCR text output for image-based detections
• Converted preview video in output/ for browser playback

Project Highlights

• Built an end-to-end ANPR pipeline using YOLO for object detection and EasyOCR for text recognition.
• Converted raw annotation formats into YOLO-ready datasets and trained a custom detection model.
• Developed a Streamlit-based inference app for image and video uploads with real-time visual outputs.
• Implemented robust media processing and browser-compatible video rendering using FFmpeg.

Future Improvements

In a production-scale next phase, I would integrate this deployed license plate detection system with CCTV streams using RTSP (Real-Time Streaming Protocol). Each frame would be processed in real time using YOLO for detection and OCR for text extraction. To optimize performance, I would add object tracking to avoid redundant OCR on the same vehicle across consecutive frames, then store detections in a database with timestamps and camera identifiers. The pipeline would be exposed through APIs and monitored through a dashboard for live events, search, and analytics.

Author

• Dev Dharmesh Patel
• Email: devdpatel190905@gmail.com
• GitHub: https://github.com/devpatel0005

License

This project is proprietary and licensed under an All Rights Reserved license. See LICENSE for full terms.