Number Plate Recognition & Helmet Detection

Two interconnected computer vision projects demonstrating end-to-end ML pipeline development: custom CRNN architecture for OCR and multi-model helmet violation detection system.

---

Project Overview

1. CRNN Number Plate Recognition - Custom OCR model built from scratch
2. Helmet Detection System - Real-time violation detection using multi-model pipeline

---

System Architecture

┌─────────────────────────────────────────────────────────────────┐
│                     HELMET DETECTION SYSTEM                      │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
                    ┌──────────────────┐
                    │   Video Input    │
                    └──────────────────┘
                              │
                              ▼
              ┌───────────────────────────────┐
              │   YOLO Object Detection       │
              │   (4 Classes)                 │
              │   • Bike                      │
              │   • Person with helmet        │
              │   • Person without helmet     │
              │   • Number plate             │
              └───────────────────────────────┘
                              │
                    ┌─────────┴─────────┐
                    ▼                   ▼
        ┌──────────────────┐  ┌──────────────────┐
        │  DeepSort        │  │  Violation       │
        │  Bike Tracking   │  │  Detection       │
        │  (Unique IDs)    │  │  (5-frame rule)  │
        └──────────────────┘  └──────────────────┘
                    │                   │
                    └─────────┬─────────┘
                              ▼
                    ┌──────────────────┐
                    │  Number Plate    │
                    │  Detected?       │
                    └──────────────────┘
                              │ Yes
                              ▼
              ┌───────────────────────────────┐
              │   CRNN OCR Model              │
              │   (Custom Architecture)       │
              └───────────────────────────────┘
                              │
                              ▼
                    ┌──────────────────┐
                    │  Output          │
                    │  • Annotated     │
                    │    Video         │
                    │  • Evidence      │
                    │    Images        │
                    │  • Plate Text    │
                    └──────────────────┘

---

CRNN Architecture

Input Image (3, 100, 200)
         │
         ▼
┌─────────────────┐
│   CNN Blocks    │  ← Feature Extraction
│   3 → 64 → 128  │     • 3 Conv blocks
│   → 256         │     • BatchNorm + ReLU
└─────────────────┘     • MaxPooling
         │
         ▼
┌─────────────────┐
│ Reshape to      │  ← Prepare for sequence
│ (B, T, C*H)     │     
└─────────────────┘
         │
         ▼
┌─────────────────┐
│  FC Layer       │  ← Dimension reduction
│  256*H → 128    │     + Dropout (0.5)
└─────────────────┘
         │
         ▼
┌─────────────────┐
│ Bidirectional   │  ← Sequence modeling
│ LSTM (2 layers) │     • Hidden: 128
│                 │     • Dropout: 0.3
└─────────────────┘
         │
         ▼
┌─────────────────┐
│  Classifier     │  ← Character prediction
│  256 → 63 chars │     (a-z, A-Z, 0-9 + blank)
└─────────────────┘
         │
         ▼
   CTC Decoding
         │
         ▼
   Output Text

---

Projects

1. CRNN Number Plate Recognition (Experimental)

Custom CRNN architecture for number plate OCR.

Features:

• Custom CNN-RNN architecture
• CTC loss for variable-length sequences
• Data augmentation
• ONNX export for deployment

Details:

• Dataset: 456 clean number plate images
• Architecture: 3 CNN blocks + Bi-LSTM + CTC
• Input: 100x200 RGB images
• Output: Alphanumeric text (a-z, A-Z, 0-9)

Results:

• Tested on blurred images
• 56% character accuracy on blur (trained on clean data)
• Demonstrates domain adaptation challenge

Prediction Results:

Sample predictions on test images

Files:

CRNN/
├── model.py         
├── train.py         
├── test.py          
├── export_onnx.py   
└── requirements.txt

Usage:

python train.py    # Train model
python test.py     # Test model
python export_onnx.py  # Export to ONNX

---

2. Helmet Detection System

Multi-model pipeline for automated helmet violation detection.

Features:

• Real-time object detection (4 classes)
• Multi-object tracking
• Violation confirmation (5-frame threshold)
• Evidence collection
• Number plate OCR

Details:

• Detection: YOLO26
• Tracking: DeepSort
• OCR: Custom CRNN
• Classes: Bike, person with/without helmet, number plate

Demo Videos:

Multiple Vehicle Tracking:

DeepSort tracking multiple bikes with unique IDs

output-video-2-al3lbu93-svhofxa2_j1TNeZ7h.1.mp4

output-video-3-23ewmukg_czRNvTTU.mp4

Violation Detection:

Person without helmet detection and tracking

output-video-2-g9jwo6yq-9hdfn38z_uhC742mN.1.mp4

Number Plate OCR:

Real-time number plate recognition

output-video_mvkHB3hx.mp4

System Output:

Evidence collection: bike, person, and number plate crops

Video.Project.2.mp4

Files:

Helmet-Detection/
├── main.py
├── requirements.txt
└── USAGE.md

Usage:

python main.py

Output:

track_output/HH-MM-SS/
├── bike_1_bike.jpg               
├── bike_1_person_no_helmet.jpg   
└── bike_1_number_plate.jpg       

output_video.mp4  # Annotated video with detections

---

Results & Learnings

What Worked:

• Custom CRNN learns clean data effectively
• Multi-model pipeline integration
• DeepSort maintains bike IDs across frames
• ONNX deployment ready

Challenges:

• Domain shift (clean to blur) reduces accuracy
• Small dataset (456 images) limits generalization
• Person-to-bike association needs spatial logic

Learnings:

• CTC loss for variable-length sequences
• Data augmentation strategies
• Multi-model integration
• Domain adaptation importance
• Model deployment (ONNX)

---

Tech Stack

Category	Technologies
Deep Learning	PyTorch, ONNX Runtime, Ultralytics YOLO
Computer Vision	OpenCV, PIL, torchvision
Tracking	DeepSort
ML Tools	scikit-learn, NumPy, matplotlib

---

Future Improvements

CRNN:

• Train on blurred data for robustness
• Larger dataset (1000+ images)
• Attention mechanisms
• Confidence scores

Helmet Detection:

• Spatial association (distance-based)
• Real-time optimization
• Multi-bike improvements

---

Installation

# Clone repository
git clone https://github.com/Pooja-Vachhad/Helmet_Detection.git

# Install CRNN dependencies
cd CRNN
pip install -r requirements.txt

# Install Helmet Detection dependencies
cd Helmet-Detection
pip install -r requirements.txt

---

License

MIT License

---

Note: Experimental project demonstrating CV concepts. Helmet detection uses simplified person-to-bike association for proof-of-concept.