ThinkAHead

Motorcycle Helmet Violation & Overloading Detection System

A real-time computer vision system that detects motorcycle safety violations including helmet non-compliance and overloading (≥3 riders). Built for CS 176 (Computer Vision) Final Project.

---

Table of Contents

• Installation
• Usage Guide
• Model Details
• Results
• Limitations
• Future Improvements
• Project Structure
• Authors

---

Installation

Prerequisites

• Python 3.10+

Step 1: Clone the Repository

git clone https://github.com/yourusername/thinkahead.git
cd thinkahead

Step 2: Create Conda Environment

conda create -n thinkahead python=3.10
conda activate thinkahead

Step 3: Install Dependencies

pip install -r requirements.txt

---

Usage Guide

Web Interface (Streamlit)

streamlit run src/app_streamlit.py

Then open http://localhost:8501 in your browser.

Features:

• Upload images or videos
• Adjust detection confidence threshold
• View real-time statistics
• Download violation logs as CSV

---

Model Details

Architecture

• Base Model: YOLOv8m (Medium)
• Parameters: 25.9M
• Input Size: 640×640
• Framework: Ultralytics + PyTorch

Detection Classes

Class ID	Class Name	Description
0	motorcycle	Two-wheeled motor vehicle
1	rider	Person on motorcycle
2	helmet	Person wearing helmet
3	no_helmet	Person without helmet
4	license_plate	Vehicle registration plate

Training Configuration

Parameter	Value
Epochs	100
Batch Size	8
Image Size	640×640
Optimizer	AdamW
Learning Rate	0.001 → 0.00001
Augmentation	Mosaic, HSV, Flip, Scale

Dataset

Split	Images
Train	1,798
Validation	224
Test	226
Total	2,248

Sources:

• Roboflow Helmet-and-Number-Plate dataset
• Roboflow Triple Riding Detection dataset
• Kaggle Helmet Detection dataset

Post-Processing Pipeline

YOLO Detection → Geometric Grouping → Violation Logic → OCR (optional)

1. Geometric Grouping: Associates riders with motorcycles based on bounding box overlap and proximity
2. Helmet Detection: Checks if helmet bbox overlaps upper portion of rider bbox
3. Overload Detection: Flags motorcycles with ≥3 associated riders
4. OCR Module: Extracts license plate text for violations (using EasyOCR)

---

Results

Overall Performance

Metric	Value
mAP@50	76.1%
mAP@50-95	39.3%
Precision	74.9%
Recall	77.2%

Per-Class Performance (AP@50)

Class	AP@50	Notes
license_plate	99.5%	Excellent detection
helmet	84.5%	Strong performance
motorcycle	80.2%	Reliable detection
no_helmet	59.6%	Moderate - challenging class
rider	27.1%	Low - often subsumed by helmet/no_helmet

Confusion Matrix Analysis

• Motorcycle: 85% correct classification
• Helmet: 87% correct classification
• No_helmet: 63% correct (some confusion with background)
• Rider: 29% correct (often classified as background)

---

Limitations

Detection Limitations

1. False positives in dense traffic - Crowded scenes may cause incorrect rider-motorcycle associations
2. Extreme viewing angles - Side/rear views may miss helmet detection
3. Nighttime performance - Depends heavily on lighting conditions
4. Rider occlusion - Overloading detection fails when riders occlude each other
5. Small objects - Distant motorcycles may not be detected reliably

Dataset Limitations

1. Class imbalance - "Rider" class underrepresented relative to helmet/no_helmet
2. Geographic bias - Primarily South Asian traffic scenarios
3. License plate formats - Optimized for certain plate styles

Technical Limitations

1. VRAM requirements - Minimum 4GB for inference
2. Real-time constraints - May require frame skipping on lower-end hardware
3. OCR accuracy - License plate reading affected by motion blur and resolution

---

🔮 Future Improvements

• Add temporal tracking (DeepSORT) for consistent ID assignment across frames
• Implement confidence calibration for violation alerts
• Add support for more license plate formats
• Nighttime enhancement using IR cameras or low-light models
• Multi-camera support for wider coverage
• Database integration for violation logging
• Expand to other violations (wrong-way driving, lane filtering)
• Edge deployment (Jetson Nano, Raspberry Pi)
• Integration with traffic management systems

---

References

• Ultralytics YOLOv8
• EasyOCR
• Streamlit

Datasets

• Roboflow Universe - Helmet and Number Plate Detection
• Roboflow Universe - Triple Riding Detection
• Kaggle - Helmet Detection Dataset

---

Authors

Noval & Sacramento

CS 176 - Computer Vision
University of the Philippines Diliman
2nd Semester, AY 2024-2025

---

📄 License

This project is for educational purposes as part of CS 176 coursework.

---